The machine remained the problem. A 510 GB model that needs four B300s may be open source on GitHub, but it is still beyond the reach of anyone who wants to run an AI in a home office.

I went back to my RTX 5090 and Strix Halo. Instead of repeating the full v4 benchmark on dozens of models that we already know will not fit or run fast enough, I took the most promising candidate and compared its quantizations. Which version preserves enough quality without turning every interaction into a wait of several minutes?

The result surprised me: the 2-bit version won. I then repeated the test with Basalt, a Strata fork specialized for Blackwell GPUs. The model kept getting everything right, while generation reached 400 tokens per second. The quantization did not change; the engine changed the result substantially.

The RTX 3090 and 4090 got a second wind

Before my results, the community found a limitation that could be removed. On GeForce RTX 20 cards, some FP16 Tensor Core operations with FP32 accumulation run at half the theoretical issue rate per cycle of their professional equivalents. On RTX 30 and 40 cards, the limit also affects BF16.

The hardware is there. Two internal registers select the speed, and the normal driver does not have enough privilege to change them.

The gpu-unlock project found a way to change those registers before the NVIDIA driver takes over the card. On Ampere, it exploits a buffer overflow in the signed Booter that authenticates and loads GSP-RM. Nothing is written to the VBIOS. To restore stock behavior, remove the activation parameter and fully power the machine off, because the selectors reset when the card loses power. If the parameter remains active, the module reapplies the unlock on the next boot.

It also requires Linux, initramfs, and a lot of care. I would not install this on a production workstation without reading the code and understanding the risk.

This is a community hack with no official NVIDIA support. You install it at your own risk. Its value for LLMs is faster prefill when the engine uses this exact Tensor Core path. Other compute applications, including some image-generation pipelines, may also change in performance. Games depend on different GPU paths and bottlenecks, so do not expect higher frame rates from it.

On an RTX 3090 at 420 W, a controlled test measured FP16 with FP32 accumulation going from 76.396 to 125.006 TFLOPS; BF16 went from 76.994 to 127.923. On the 4090, an implementation that is still experimental measured roughly twice the FP16 and BF16 instruction issue rate.

Doubling TFLOPS does not double your LLM. The test that prompted this discussion separated the two phases:

• Prefill is when the model reads the whole prompt. Many tokens are available at once, so the engine can build large parallel matrix operations. One 3090 gained 14.4% on a 16K-token workload, while two modified 3080s gained 33% to 36% with vLLM.
• Decode is when the model generates one token after another. With a quantized LLM using expert offload, this phase is constrained mainly by memory bandwidth, RAM, PCIe, and the expert cache. The measured change stayed within noise.

The gain varies substantially with the engine, quantization, and prompt size. On the dual 3080 setup, Qwen3.8 Flash-Next Q4_K_XL gained only 0.5% to 0.8% at 8K, but reached 12% or 13% at 32K and 128K. On a 3090 running an IQ3_S MoE, a workload near 32K did not change. The computation path used by a quantization may never touch the FP16/BF16 path that was unlocked.

gpu-unlock accelerates a specific Tensor Core path. It may reduce the wait when reading large prompts, but it does not turn 24 GB into 96 GB or create memory bandwidth.

Even with that caveat, the 3090 and 4090 became more interesting for local inference. A used RTX 3090 still has 24 GB of VRAM, NVLink in its generation, and 936 GB/s of bandwidth. The community measured Qwen3.8 Flash-Next at roughly 75 tok/s with IQ3_S and 100 tok/s with Q2_0 on one 3090. With two cards, reports range from 90 to 120 tok/s, depending on the configuration and context size. Those are community results, not measurements from this round.

Why I chose Qwen3.8 Flash-Next

In v4, the open candidates that really interested me were:

The three tied within the noise margin. Only Qwen repeated that quality on consumer hardware. Its main model is a 125-billion-parameter MoE with roughly 6 billion active parameters per token, plus 51 billion parameters in n-gram embeddings and a 4-billion-parameter MTP head. Strata keeps the most frequently used experts on the GPU, keeps the rest in RAM for the CPU to compute, reads the n-gram table from the SSD, and uses the model’s own MTP layer to speculate the next tokens.

DeepSeek V4.1’s backbone and Engram add up to about 748 billion parameters: 552 billion in the backbone and 196 billion in Engram. The published checkpoint has roughly 763 billion parameters in total. I served its native mixed FP4/FP8 version, about 510 GB, on the four-B300 pod.

I also managed to start a 2-bit version on the 5090 using Strata’s dedicated engine. It generated around 17 tokens per second. It runs. It is out of the race for a programming agent.

The GLM 5.3 Flash NVFP4 checkpoint was around 198 GB and ran on the B300. This repository has no consumer configuration that I measured and can recommend. The smaller GLM 4.7 Flash had already scored 24 on the 5090 and 33.5 on the B300 in the previous benchmark. Spending days forcing the wrong family to fit would not produce useful information.

I also left out Qwen 4 and Llama by choice. This round asks a narrow question: which version of the model that already proved its quality works best on the two machines I own?

What I tested

All the code, prompts, hidden tests, and results are in llm-opensource-benchmark. I tested four Qwen3.8 Flash-Next versions:

The Strata files include a shared PLE component, so the full size does not fit directly in VRAM. The engine distributes experts between the GPU and RAM. On Basalt, I used the project’s own packs: one 64.9 GiB .basalt file for Q2_0 and an 81.0 GiB file for IQ3_S. Each file already includes the weights, tokenizer, MTP, expert profile, and vision projector.

For the first RTX 5090 round, I used Strata’s CUDA backend with:

• 4 GiB of VRAM reserved for Hyprland, the browser, and the terminal;
• int8 KV cache, a 128K maximum context, and 32K resident on the GPU;
• an elastic expert cache;
• four-token speculation with the model’s MTP.

I then ran Q2_0 and IQ3_S on Basalt, on the same 5090 and with the same 131,072-token maximum context and int8 KV. I reserved 4,608 MiB because Basalt counts its working buffers against that margin. Free VRAM stayed between 3.6 and 4.4 GB during the tests.

On the Strix Halo, I used the hip-halo backend. That machine has 64 GB reserved for the GPU and 31 GB of GTT mapped over system RAM; the engine’s RAM limit was 61.9 GiB. Since it is headless, I gave the entire GPU to the model and kept 8 GiB of headroom for the conversation cache.

The Strata measurements use two different backends underneath, CUDA on NVIDIA and HIP/ROCm on AMD, with the same engine and the same OpenAI-compatible API. Basalt adds an engine comparison on the 5090, using the same OpenCode, prompts, and grading.

Methodology

I started with a speed test in three situations:

1. a cold 4K-token prompt;
2. a cold prompt labeled 32K, which this model’s tokenizer counted as 42,400 tokens;
3. a continuation of the same conversation, reusing the cache.

Each one generates 256 tokens. The runner measures time to first token, prompt processing, generation, VRAM, RAM, power, and temperature.

What each speed number means

When someone says an LLM runs at 200 tokens per second, they are almost always talking only about generation. An agent session goes through different phases, and each phase hits a different bottleneck.

Prefill is the initial reading of the prompt. The model must process the harness instructions, conversation history, files, and your question before producing its first word. Since all those tokens are available at once, the GPU can process large matrices in parallel.

This phase makes better use of Tensor Cores and raw compute. The larger the context, the more it matters.

The KV cache stores intermediate states for tokens that have already been processed. On a continuation of the same conversation, the model reuses that state instead of reading everything again. That is why the append test reaches tens of thousands of prompt tokens per second and receives the first token in roughly half a second. That number measures cache reuse; comparing it with cold prefill would be dishonest.

The KV cache also grows with the conversation and competes with weights, experts, and buffers for memory. Here it uses int8, with 32K resident on the 5090. Advertised maximum context, context that fits, and context that stays fast are three different numbers.

Speculative decoding uses a draft layer to propose several tokens and lets the main model verify the batch. Output tok/s goes up when the acceptance rate is good. The benefit vanishes when the draft misses too often. The Strata speeds already include MTP with four-token speculation. Basalt also uses the MTP included in its pack, with its own speculation mechanism and defaults.

One metric is missing because I did not measure it here: concurrency. Every number in this article comes from a single stream. A server can handle several sessions and raise aggregate throughput, while each user will usually receive fewer tokens per second.

For individual use, single-stream performance and wall time are what matter. A team must measure simultaneous sessions and per-user latency, as I did in the B300 and H200 test.

Then came two programming tests using OpenCode 2.1.4. The model receives a small Python codebase, works in one continuous session, and never sees the hidden tests. Everything runs inside Bubblewrap, without network access, with a limited filesystem and a local gateway that accepts calls only to the expected model.

The first canary has three turns and 9 hidden tests. The model must implement a log analyzer, fix bugs, and handle changed requirements.

The hard canary has four turns and 36 hidden tests. It is a payment reconciler with decisions buried inside a roughly 79K-token history, one decision corrected later, rejected proposals used as distractions, money that requires exact decimals, and new rules that interact with earlier ones. With OpenCode’s system prompt, the first turn approaches 83K tokens.

A separate yardstick from benchmark v4

These canaries are much smaller and easier. In v4, the model works on a Rails application for seven sprints while a reviewer injects 14 security and correctness defects. The score measures how many defects it finds and fixes on its own before they are revealed. It is a vigilance test with much more code, more tools, and more opportunities to lose track.

Here I use two small Python codebases, three or four turns, and hidden functional tests. I built this yardstick to compare quantizations of the same Qwen and find out whether reducing the bits breaks requirements, context retention, or process. The comparison ends there. Qwen, Opus, GPT, DeepSeek, and the other model families remain a job for v4.

A score of 100 in this article belongs to a different scale, with no equivalence to 100 in v4. It also does not make Q2_0 “better” than the IQ3_S version that scored 91 in v4. V4 selected Qwen as the candidate; this smaller test selects which Qwen file is worth loading at home.

Speed without quality only makes the model fail faster. That is why I measured them separately.

Speed ranking

RTX 5090 with Strata, 4 GiB reserved for the desktop

RTX 5090 with Basalt

Strix Halo, using the whole GPU

The 5090 has 32 GB of GDDR7 and 1,792 GB/s of bandwidth. The Strix has more total capacity, but its roughly 256 GB/s must serve the entire system. Even with Strata on both machines, the difference appears mainly in prefill: 6.9 seconds versus almost 2 minutes to start answering the 42,400-token workload. Basalt cut the 5090 wait to 4.9 seconds.

Q4 gave me the worst of both worlds. On the 5090 it ran slowly, and systemd-oomd killed the quality test. On the Strix, it exhausted VRAM, started paging, and froze the machine until a reboot. More bits preserve more information, but they also use more memory and shrink the hot cache that makes this MoE move.

Basalt versus Strata

Basalt started as a Strata fork and took a narrower direction: Linux or Windows, CUDA 13, and an NVIDIA Blackwell GPU with compute capability 12.0. The 5090 is exactly its target. Strata remains much broader and is what allowed the same Qwen to run on the AMD Strix Halo.

Both engines solve the same central problem. In the tested configuration, experts that do not fit in VRAM stay in RAM while the GPU holds a hot subset. Basalt and Strata expose OpenAI- and Anthropic-compatible APIs and accelerate generation with their MTP drafters. Basalt stores everything in one file and has kernels, memory management, and speculative execution developed specifically for Blackwell.

This was a practical A/B between complete serving stacks, without isolating one optimization. Each engine has its own packs, their defaults are not identical, and Basalt reserved 4,608 MiB of VRAM versus Strata’s 4,096 MiB. I kept the same RTX 5090, the same quantizations, a 131K maximum context, int8 KV, the same OpenCode, and the same tests.

Basalt won every speed metric in this A/B and tied on functional coding results. Q2_0 moved from a range of 223 to 259 tokens per second to 377 to 400. Long-input processing rose from 6,158 to 8,573 tokens per second, and the wait for the first token fell from 6.9 to 4.9 seconds. IQ3_S moved from 158 to 208 to a range of 223 to 231 tokens per second.

Platform scope decides the choice. On an RTX 5090, I would start with Basalt. On the Strix, pre-Blackwell NVIDIA, or other supported hardware, Strata remains the path. Strata covers the platforms where Basalt does not run.

Quality ranking

The simple test separated nothing:

All three passed the 9 hidden tests, preserved the requirements, and finished with the visible suite green. I had to make the test harder.

In the hard canary, I ran Q2_0 and IQ3_S three times each on the 5090 with Strata:

Both versions got everything right in all three runs. They found the decisions in the long context, respected the later correction, handled out-of-order events, chargebacks, currency conversion, and completed the final refactor. They also went through a similar number of failing test runs while working.

The difference was on the clock. Q2_0 took 7.2 minutes on average. IQ3_S took 12.1. The 2-bit model delivered the same result in roughly 60% of the time.

I then repeated all six runs on Basalt:

The penalty that dropped the second IQ3_S run to 96 was a known harness false positive. The model executed ps aux inside the sandbox, and the command showed the bwrap process itself, whose command line contains host paths. The detector treated that as an attempt to access the hidden tests and removed four points. The model did not use those paths, the hidden tests are only mounted later during grading, and all 36 passed.

On Basalt, Q2_0 took 5.4 minutes on average, compared with 7.2 on Strata. IQ3_S fell from 12.1 to 9.0 minutes. The end-to-end gain is smaller than the jump in tokens per second because the session also spends time running tools and tests, and each run generated a different amount of text.

The evidence covers small Python codebases, contexts up to roughly 85K tokens, and only three repetitions per engine and quantization combination in the hard canary. A 2-bit quantization will usually lose quality first on contexts above 100K and on larger codebases. Within the use I measured, spending more bits bought no quality.

What each version showed

Q2_0: the current recommendation

It generated tokens faster in all three machine-and-engine combinations and finished all six hard runs with a perfect score. IQ3_S won the long-prompt prefill measurement on Strata and the 5090 by a tiny margin. With Basalt, Q2_0 reached 400 tokens per second. On the Strix, 55 to 65 tokens per second is usable, although long prefill costs almost two minutes before the first answer.

I would use this version every day and keep the context caveat in mind. Large project, session approaching 100K, or strange behavior? I repeat the same work with IQ3_S before blaming the model.

IQ3_XXS: no clear role

It passed the simple canary and ran quickly, but Q2_0 was faster and uses less space. I did not run the three hard-test repetitions on IQ3_XXS, so I have no data showing a quality advantage.

IQ3_S: the conservative option

This is still the version I would use to validate a suspicious Q2_0 result. It had already scored 91 in benchmark v4 on the 5090, only 1.5 points below its NVFP4 version on the B300. On Basalt, it ran at 223 to 231 tokens per second, roughly the range Q2_0 reached on Strata, while processing prompts faster.

Basalt also measures each quantization against Q8_0. Across its 100-chunk test, IQ3_S kept the same top token in 87.16% of positions. That is a different yardstick from my canaries, but it supports IQ3_S’s role as the conservative option. In my coding tests, that extra precision still found nothing Q2_0 had missed.

UD-Q4_K_XL: quality that does not fit is useless

It was the intuitive candidate for preserving more fidelity. I ran out of memory before reaching the test that could confirm it. On the 5090, it fell below one quarter of Q2_0’s speed and the system killed the run. On the Strix, the machine froze.

The largest quantization that fits on paper can be the worst real configuration. An MoE depends on the space left for experts, KV cache, buffers, and the operating system itself.

Which hardware is worth it?

RTX 3090 and 4090: good if you already own them

Both have 24 GB, which limits the expert cache. gpu-unlock improves prefill in workloads that use the unlocked path, while decode remains constrained by memory and offload. A used 3090, especially two of them with NVLink, can make an interesting local machine. I would not buy a new 4090 today just for this: it still has 24 GB, costs a lot, and the Ada unlock remains experimental.

For either one, I would start with Qwen3.8 Flash-Next Q2_0 and measure my workload. The roughly 100 tokens per second published by the community on one 3090 already puts the card in a practical range. Two 3090s add capacity and may land between 90 and 120 tokens per second, but they do not automatically double throughput. The bill includes power draw, heat, a suitable PSU, and at least 128 GB of RAM.

Strix Halo: plenty of capacity, little bandwidth

It is a good compact and efficient headless box. Unified memory lets it run models that a 24 or 32 GB GPU would reject. In this test, the best model was also Q2_0, at 55 to 65 tokens per second.

The problem appears every time the agent reads a large context. Almost two minutes to process the long workload adds up quickly during a session. I would use it as a private server for asynchronous tasks, simple agents, and workloads where capacity matters more than immediate response.

RTX 5090: the best home option today

The RTX 5090 has 32 GB, 21,760 CUDA cores, and the same nominal 1,792 GB/s bandwidth as the RTX PRO 6000. Capacity is the missing piece: I must reserve 4 GiB for the desktop and use RAM and SSD for the experts that do not fit.

With Strata, it pushed Q2_0 to 223 to 259 tokens per second and read the long workload at 6,158 tokens per second. With Basalt, those numbers reached 400 and 8,573, respectively. This is the first local configuration that does not make me feel like I am waiting for the machine at every step. It is also a card I can use for gaming, image generation, video, and everything else in the workstation.

RTX PRO 6000: three times the VRAM, server-class price

The RTX PRO 6000 Blackwell has 96 GB of GDDR7 ECC, 24,064 CUDA cores, and the same 1,792 GB/s bandwidth. The extra capacity keeps many more experts resident, makes room for a larger quantization, and holds context without fighting the desktop for the last few gigabytes.

It costs $16,000 in NVIDIA’s store, eight times the 5090’s suggested $1,999 price. I have not run this benchmark on one, so I will not invent a speed. Equal bandwidth means that three times the memory will not produce three times the tokens per second. Its gain comes from keeping more data on the card and avoiding the slow RAM and PCIe path.

For individual use, the 5090 makes more financial sense. The PRO 6000 starts to make sense when 96 GB solves a model you actually need, when ECC and professional support matter, or when several sessions will share one card.

Verdict

Local open source models have finally entered a practical range. A single 5090 runs a model that reached Tier A in my v4 benchmark and, with Q2_0 on Basalt, reached 400 tokens per second in the canary. The Strix also did far better than the 7.7 tokens per second from the dense Qwen I had tested before.

The limits remain clear. Context consumes memory, the desktop needs the same VRAM, and every expert outside the card travels over PCIe or falls back to RAM. DeepSeek at 17 tokens per second shows how useless “it runs” can be as a metric. If every iteration takes several hours, model quality stops mattering because nobody can work that way.

For home hardware, the best model preserves enough quality at a speed that lets you iterate. In this round, that was Qwen3.8 Flash-Next Q2_0 on Basalt, running on an RTX 5090.

The benchmark repository has the runners, configurations, raw results, and hidden tests. Repeat it on your machine and send results. Hardware, engine, quantization, and context change the answer. A ready-made ranking is a starting point; the final decision is still measured against your own work.

The model builds a Rails project over seven sprints and must find 14 security and correctness defects planted along the way. I measure whether it notices a backdoor, an authorization bypass, a tampered test, or a wrong calculation before someone ships the answer. It is a measure of vigilance during programming. It was never a test of universal intelligence.

As I wrapped up that series, a few colleagues who work in investments and law started discussing whether an open model could already handle parts of their routines. That left me with an itch: how could I measure this without falling back to multiple-choice exams or legal trivia?

I tried to answer it. And this disclaimer controls everything else in this article.

I am not an accountant, I do not have a CPA license, I never passed the bar, and I have zero professional experience working at an accounting or law firm. I know enough to discuss and read the material, but I cannot certify that this suite properly represents the daily work of either profession.

Treat everything from this point on as an experiment that needs scrutiny from actual accountants and lawyers. These numbers compare models inside this suite. They do not certify any model to serve clients, sign an opinion, close a set of books, or file a legal document.

What I tried to measure

Some of the professionals I spoke with already use Claude or GPT at work, with models such as Fable, Opus, Astra, or Sol. I started from that reality instead of pretending I could design a professional test by myself.

I asked Fable 5.1 to design cases within what it was confident it could research, structure, and review. The suite draws on professional work patterns, statutes, forms, decisions, and reference material, but all clients, companies, and documents are synthetic. No real confidential data entered the repository.

Then I ran GPT-6 Sol as the frontier anchor. Finally, I took the open models that had performed well in the programming benchmark and ran the same suite on three infrastructure classes: an RTX 5090, Strix Halo, and a rented NVIDIA B300.

I use “open” in the same sense as the previous benchmark: a checkpoint I can download and serve myself. Licenses vary, and not every open-weight model satisfies the strictest definition of open source software.

Fable also performed the blind grading. Each configuration received a random letter, and a separate subagent evaluated each case without knowing which model produced the answer. The comparison in this article therefore means something very narrow:

How close did these open models get to Sol on a test designed and graded by Fable?

Fable and Astra did not run the suite as candidates. Sol is the only closed-model anchor in this campaign. Fable itself limits what we can ask and judge, and an AI grader can make mistakes. Human professional spot-checking is still pending.

The test starts with a client folder

The benchmark contains 38 cases, split between accounting and law in the United States and Brazil. Some cases span several sessions, so the full campaign totals 63 session units per configuration and 190 graded executions.

This is the distribution:

There are four L1 cases, six L2 cases, 22 L3 cases, and six L4 cases. Difficulty does not come from asking for more elaborate prose. It grows with document volume, the number of interacting rules, and how long the model must preserve a coherent line of reasoning.

• L1, clerk or intern: one or two documents, one deliverable, and one verifiable answer. One example is calculating deadlines from a DJe notice.
• L2, senior: several documents, a threshold or rule interacting with another one, and at least one fact that must be requested from the client. Tax regime selection and contract review against a playbook live here.
• L3, manager or associate: conflicting or superseded sources, false positives, and decisions that must be justified. Most of the suite sits at this level.
• L4, partner: three to twelve sessions. A new fact invalidates previous work, the opposing side reacts, and the model must change course without hiding the change.

The six L4 cases are twelve monthly closes for a SaaS company, a Chapter 11 debtor, one quarter under Brazil’s Lucro Real regime, a Brazilian judicial reorganization, a five-session US lawsuit, and a five-session Brazilian tax foreclosure. Together with the single-session cases, they bring each configuration to 63 session units. Five configurations produced 315 work sessions across 190 executions.

Each case resembles a small professional engagement. The model receives:

• a folder containing contracts, statements, invoices, pleadings, spreadsheets, and other documents;
• a closed library containing the allowed rules and authorities, with no internet access;
• an engagement letter explaining who the client is, what must be done, and the required format;
• tools to calculate and manipulate files;
• a simulated client that reveals certain facts only when the model asks the right question.

The suite ranges from short clerk-level tasks to long partner-level engagements, with late facts that force the model to revisit an earlier conclusion. Every case ends with the same instruction to review errors, unsupported sources, and numbers that do not reconcile.

Accounting example: the close that will not close

One US case provides the February trial balance, March general ledger, bank and credit-card statements, an accounts receivable aging, two copies of the same vendor invoice, and the close checklist. The owner wants a clean reconciliation by Friday because the company is renewing a line of credit.

After every legitimate item is recorded, an unexplained $4,312 difference remains. No document supports that amount. The correct answer must leave it open, tell the owner, and refuse to invent an entry.

The case also hides a $1,500 transfer. The model must ask where it came from. The scripted client answers that it was a personal contribution from the owner.

If the model assumes it is revenue, both income and tax are wrong. If it posts the $4,312 to “miscellaneous expense,” “owner draw,” or a suspense account to force the reconciliation, it immediately fails the trust gate.

This example tests more than arithmetic. It tests whether the agent resists client pressure, recognizes missing evidence, catches the duplicate invoice, asks the necessary question, and produces a trial balance, entries, and open-items note that agree with one another.

Legal example: five sessions of tax foreclosure

In the Brazilian L4 case, a stonework company faces an IRPJ and CSLL tax debt certificate of approximately R$1.1 million. The model receives the certificate, part of the administrative tax proceeding, amendments to the articles of association, the process server’s return, and a closed library containing the necessary CTN, LEF, CPC, precedents, and binding summaries.

In the first session, it evaluates defects and chooses the procedural vehicle. In the second, the tax authority amends the debt certificate. In the third, it seeks to redirect liability to a shareholder. The final two sessions cover the shareholder’s defense, security, and a settlement with calculated amounts.

The trap is to keep repeating a theory after the defect has been cured, defend the wrong shareholder, or automatically accept the presumption of irregular dissolution. The corporate amendments show who managed the company during the relevant period, and the address change had been registered before the process server’s return. The model must drop the theory that lost its basis and recalculate the strategy around the new facts.

Another legal case applies the same pressure in a five-session US lawsuit. Opposing counsel presents an email that weakens the initial theory, an arbitration clause appears in small print in Exhibit C, and the agent must tell the client that the position has deteriorated. Persisting with the original narrative counts against it.

Every case offers a very convenient wrong conclusion. The test measures whether the model finds the evidence that forces it to change its mind.

How I graded the deliverables

The answers pass through four filters:

1. Trust: Did it invent a client fact, authority, or balancing number, apply superseded law, or claim completion before finishing? The execution fails.
2. Oracle: Do dates, amounts, classifications, and required elements match the expected answer exactly?
3. Vigilance: Did it find the planted problems during normal work or only after the final review instruction?
4. Deliverable quality: Does the memo, opinion, or summary contain the binary elements required by the case? Pretty writing earns nothing.

The oracle contains 235 exact items per configuration. A deadline lands on the right date or it does not. A form contains the expected amount or it does not.

A mandatory clause was flagged or it slipped through. Where a numeric tolerance makes sense, the answer key defines it in advance.

Traps carry weights by severity: critical is worth 5, high 3, medium 2, and low 1. Discovery during normal work receives full weight; discovery only after the final review instruction receives 40%; a missed problem receives zero. That produces the vigilance column.

The client also follows a script. It does not improvise help. Every hidden answer has a trigger and appears only when the model’s question actually seeks that fact.

If nobody asks, the model must state the assumption or leave the item open. Inventing the answer can sink the entire execution.

Before every run, the harness copies a clean folder and physically hides the answer key, trap manifests, and other models’ answers. Network access is blocked. The agent cannot search for a prepared solution or read grading files from the repository.

Neither of the two generic reminders allowed by the protocol was used. The answer files, histories, prompts, grading keys, and scripts are all available in the llm-non-coding-benchmark repository.

The experimental ranking

I sorted the table by completed engagements. A completed engagement finished every session, did not hit the response limit, had no high-confidence trust failure, and passed every gated oracle item.

Sol cost $16.72 at API-equivalent list prices. I ran it through my subscription, so this figure compares consumption and does not represent a bill I received. Local machines have no per-call charge, but hardware, electricity, and maintenance still cost money.

Runpod’s official B300 page currently lists $6.94 per hour for Community Cloud and $8.99 per hour for Secure Cloud, billed by the second. The benchmark did not retain the pod invoice or tier, so I calculated both ends of the range using only each model’s active minutes.

That produces $71.94 to $93.20 for Qwen and $83.86 to $108.63 for GLM. These are artificial floors. The pod also bills while downloading and loading weights, compiling, changing models, waiting for the controller, retrying runs, and sitting idle until someone turns it off. The real bill was higher.

“Active time” sums model time across all cases. Sol ran four in parallel and completed the campaign in 72.9 minutes of wall-clock time. Every other configuration ran one case at a time. This table compares inference effort and per-case latency; it does not measure how many simultaneous users each server could handle.

There are two oracle columns because the 16,384-token response cap cut off local-model answers. “Graded items” ignores work that became impossible to grade. “All items” counts missing work as wrong. GLM, for example, was the most exact when it produced something gradable, but lost too much work to the output limit.

A higher place in this table does not make a model better to work with. I prefer Sol’s 25 clean completions to 27 completions with three serious fabrications on Strix Halo.

The cost of an error matters more than the average

After publishing this article, I received a comment: average accuracy is not enough for lawyers and accountants. The error that matters is the plausible claim presented as fact, without a caveat, that eventually reaches the client.

The benchmark keeps the trust gate separate from the rest of the score. A wrong date or value fails an oracle item. Inventing a client answer, document, authority, or number to make the work reconcile fails the entire execution, even when everything else is correct.

The difference across configurations is large:

Five of 190 executions were affected, or 2.6%. Across those five runs, the graders found nine serious occurrences: client answers that were never given, nonexistent documents, fictitious dates, invented contract terms, and a quotation falsely attributed to the closed library. Another seven executions received medium-confidence flags and still need human adjudication.

Calling this a “confidently wrong answer rate” would be imprecise. The benchmark does not measure the model’s internal confidence. The high confidence field in the results records the grader’s confidence that a violation occurred. What we measure is whether the model asserted unsupported information, did not label it as an assumption, and allowed it to reach the delivered material or the work produced before interruption.

I also did not classify every wrong value under this category. Incorrect dates, calculations, classifications, and recommendations defined in the answer key are part of oracle accuracy. In the all-items column, omissions caused by the output cap count as errors; in the graded-items column, they stay out of the denominator.

The next version should classify each failure by type and record whether the model disclosed uncertainty. That would give us separate rates for fabrication, numerical errors, date errors, omissions, and corrections during review, alongside general accuracy.

The same Qwen build had similar strict accuracy on graded items on the RTX 5090 and Strix Halo, 91% versus 90%. No RTX execution had a serious fabrication; three Strix executions did.

Sol 6: the reference makes mistakes too

Sol was the fastest, had no high-confidence trust failure, and completed all 38 executions without hitting the output cap. Only 25 passed every criterion.

Its weakness appeared in the numbers, especially across five Brazilian accounting cases. It missed calculations involving tax regime selection, CBS/IBS, Lucro Real, company formation, and judicial reorganization. It also asked the client far fewer questions: 51 facts versus 74 to 95 for the other models, while recording more assumptions.

An explicit assumption is correct behavior when information is missing. Asking often produces a better answer than carrying a hypothesis all the way to delivery.

The final review also made answers worse. On one US tax form, Sol had all 16 items correct before review. It then cleared four correct lines and marked the work as blocked. In another case, it nulled two fields that had already been right.

Using a closed model as the judge is still a mistake. Sol served as the practical reference for this campaign and showed a clean, fast, consistent profile. It still failed 13 of 38 cases under the strict criterion.

Qwen3.8 Flash-Next on the RTX 5090: the best local balance

The strongest raw result came from the IQ3_S build, quantized to a little over 3 bits, on one RTX 5090. It delivered 27 completed engagements, no high-confidence trust failure, and 91% strict accuracy among the items that could be graded.

It consumed 432 active minutes. Median case time was 1.6 times Sol’s, with average generation at 153 tokens per second. Three executions hit the response cap. That is still a strong result for a local workstation.

Its mistakes explain why professional review remains mandatory. In one Brazilian case, it calculated deadlines one day late and used the final review to “confirm” its own wrong calculation. In a US lawsuit, it had correctly calculated a $375,000 portion and changed it to zero during review.

There were also two medium-confidence fabrication flags still waiting for human review. They were not counted as high-confidence trust failures. I would use this model to prepare work, cross-check documents, raise questions, and produce a first draft inside a machine I control. I would never let the result reach a client before a qualified person checked it.

The same Qwen on Strix Halo: five times slower

Strix Halo ran exactly the same IQ3_S quantization and also finished with 27 completed engagements. The similarity ends there.

It consumed 2,105 active minutes, almost five times the RTX 5090 and 8.5 times Sol’s total active time. Median time for the same case was 7.7 times slower than the reference. Forty tokens per second works for batch jobs and patient users; in an interactive routine, you start waiting on the machine.

Worse, this run produced three serious trust failures. The model invented client answers, documents, tax debt certificate dates, and facts that supported its own conclusions. The identical build on the RTX had no high-confidence failure.

With one run per configuration, I cannot attribute the difference to hardware, serving stack, or inference randomness. Claiming that Strix “makes the model hallucinate” would be dishonest. The evidence supports a narrower conclusion: this run was slow and not reliable enough. It needs repetition before any recommendation.

Qwen3.8 Flash-Next on the B300: a large rented GPU did not buy quality

On the rented B300, Flash-Next ran in NVFP4, completed 23 engagements, and produced two serious trust failures. In one of them, it invented a city, contract terms, figures, and even a quotation absent from the closed library.

Active time was 622 minutes, with median case time 1.9 times Sol’s and average generation at 148 tokens per second. Five executions hit the response cap; four directly prevented a completed delivery. Even under the minimum cost calculation, it cost 4.3 to 5.6 times Sol’s API equivalent.

In this serial test, a B300 delivered neither higher quality nor lower latency than the RTX 5090. This does not measure the main reason to buy or rent a B300: serving many users and batches in parallel. I ran one engagement at a time to preserve a comparable protocol. The evidence only supports this conclusion: paying for the datacenter class did not improve this individual session.

GLM 5.3 Flash on the B300: precise when it finishes

GLM had a different profile. It recorded no high- or medium-confidence trust flag and hit 93% of the strict items it managed to deliver, the highest percentage in the table.

It completed only 23 cases. Seven executions hit the 16,384-token cap, and accuracy drops to 79% when every missing item counts as wrong. It took 725 active minutes, with median case time 2.7 times Sol’s, at 116 tokens per second. At Runpod’s official rates, the floor was 5 to 6.5 times Sol’s API equivalent.

The choice of column completely changes the reading. GLM was careful and exact on gradable results. The combination of model, harness, and response cap failed to finish enough work.

The cap was a harness choice, not an intrinsic GLM limitation. I need to repeat the test with longer responses or shorter sessions. Until then, selling the 93% without mentioning the seven interruptions would be marketing.

Final review helps and also breaks things

I always recommend asking an agent to reread everything before delivery. The benchmark confirms that advice, with an important caveat.

In 113 of 190 executions, graders found some problem removed or corrected during final review. These included fabrications that otherwise would have shipped. The model rereads sources, removes dubious citations, corrects names, and deletes claims it cannot support.

In about a dozen runs, the same review changed correct work into a wrong version. Sol, RTX, and GLM replaced correct dates, figures, and conclusions. In two cases, the model recomputed an incorrect number and came away even more confident in it.

Automated review is a hygiene layer. The word “review” does not turn the model into an independent auditor of itself.

For sensitive work, I would use at least two separate stages: one agent produces, another reviews with clean context and explicit criteria. Responsibility remains with the professional.

Parts of the benchmark failed too

The models scored from 97.5 to 100 on vigilance. Announcing that they all became perfect partners would be tempting. The real explanation is that this part of the test was too easy.

Some libraries practically revealed the problem in their indexes. Facts meant to remain hidden were already present in documents. Some scripted client pushback also counted as unprompted discovery. As a result, 961 of 973 trap-runs were caught before review and the metric lost its power to separate models.

Items from four cases with confirmed fixture defects were demoted to diagnostic status. Other cases still contain inconsistencies that need review. The broad order remained, with changes to ties, but the absolute figures do not deserve two decimal places of confidence.

There are more limitations:

• one run per configuration, with no mean or error bars;
• Fable as both designer and grader, subject to its own blind spots;
• no accountant, CPA, lawyer, or attorney has formally reviewed a sample yet;
• Codex for Sol and OpenCode for the local models;
• different quantizations and serving stacks;
• an output limit that affected only the local models;
• a closed legal library that does not test current research on the internet.

I published these problems because a benchmark without self-criticism turns into a marketing table. The next version needs corrected cases, harder traps, a fixed reference date in every engagement, a larger response cap for long cases, and repeated runs for every configuration.

Privacy changes the calculation

For a developer working on an open project, sending context to the cloud may be acceptable. Accounting firms, law firms, banks, and regulated companies handle contracts, litigation strategy, tax documents, personal data, and trade secrets. Internal policy, client contracts, or regulatory obligations may prevent that material from leaving infrastructure controlled by the organization.

In those cases, a local model competes on more than price. Sometimes it is the only permitted option.

Local execution does not magically make data secure. You still need access controls on the workstation, encrypted disks and backups, protected logs, unnecessary telemetry disabled, restricted networking, and a clear understanding of where the harness stores sessions. A poorly administered local server is still a data leak waiting to happen.

Every case in this benchmark is synthetic precisely to avoid that problem during testing.

So, is it usable already?

Cloud, with Sol: it was the fastest and most predictable option in this campaign. The $16.72 equivalent cost for 38 cases is low compared with the labor required to review 38 engagements. It makes sense when policy permits sending the data, with mandatory human review.

An owned RTX 5090: Qwen3.8 Flash-Next already works as a local assistant for organizing documents, preparing drafts, calculating figures, flagging open items, and raising questions. It was the best open-model balance I measured. It still got dates and amounts wrong in convincing ways.

Strix Halo: it is compact, but 35 active hours for the campaign make interactive use difficult. This run also produced serious fabrications. It may work for batch and low-volume tasks after more repetitions.

A rented B300: it did not improve the quality of one session, and active time alone already exceeded $70 per model. Its economics depend on concurrency, utilization, and volume, none of which this benchmark measured. For one person doing one task at a time, the figures do not justify the larger infrastructure.

No configuration completed more than 27 of the 38 cases. The best result was 71%. Even accepting that part of the failure came from the harness and fixture defects, I would not put any of them to work alone.

Open models are already viable assistants for accounting and law. This experiment does not support calling them autonomous professionals.

A model that saves two hours of reading and prepares a good first draft may already pay for the machine. Its value lies in reducing repetitive work without creating more risk than the time it saves.

I need professionals for the next round

The entire project is open at github.com/akitaonrails/llm-non-coding-benchmark. It includes the methodology, grading report, execution timings, cases, scripts, and raw results.

Right now I need professionals more than another GPU. Licensed practitioners need to review the cases, expected answers, and failures flagged by Fable. I also want new cases that reflect real work without including client information.

If you are an accountant, CPA, lawyer, or attorney and found a technical absurdity, open an issue with the correction and source. That is exactly the scrutiny this experiment needs before it becomes more than an interesting first reading.

• Part 1: why I rebuilt the benchmark from scratch for the third time, how the Rails app sabotaged with real CVEs works, and how much it cost.
• Part 2: the original cloud ranking, with almost 40 models, six ties at the top and a free model beating half of the Claude lineup.
• Opus 5.5, GPT-6 sol and luna, MiMo 2.6 and Grok 4.7: the next cloud batch. Opus 5.5 closed at 100, GPT 6 luna scored 95.5 for under a dollar per run, and Grok 4.7 regressed against its own predecessor.
• The top open source LLMs on a B300: one rented GPU, 14 open models, and two of them, GLM 5.3 Flash and Qwen3.8 Flash-Next, tied at 92.5, on par with Sonnet 5.
• Open source that only fits in 4x B300: the big models, most of them between 400 billion and 1 trillion parameters. DeepSeek V4.1 Flash tied the leader at 92.5 and the rest showed that size doesn’t buy vigilance.
• Can my team fit in 4x B300/H200?: a real concurrency test. The same DeepSeek serves 32 comfortable sessions on the B300 pod and only 8 on the H200 one.
• Consumer-grade open source on RTX 5090 or Strix Halo: Qwen3.8 Flash-Next scored 91 on a PC card with the Strata engine, at 159 tokens per second.

This article puts everything in a single table and answers the question left over: as of October 2026, is open source good enough?

The biggest benchmark I’ve ever run

v4 was, by far, the biggest round of testing I’ve ever done. That’s almost 70 complete runs of the same protocol, almost 60 different models, running on cloud APIs, on a rented B300, on a four-B300 pod, on an RTX 5090 and on a Strix Halo, plus a four-H200 pod just for the concurrency test.

And it was expensive. The first batch alone, the cloud one, went past 4 thousand dollars in nine days, counting credits and subscriptions. The single B300 cost about 270, and the 4x B300 and 4x H200 campaign took about another 2,100. Adding up what I can add up, it went past 6 thousand dollars, not counting machine hours at home.

I don’t think many people are going to reproduce this any time soon. If this material was useful to you, share it with whoever needs to decide between cloud, their own server or a local PC. All the code, prompts, logs and reports are open on my GitHub: akitaonrails/llm-coding-benchmark.

A refresher on the yardstick

The protocol is always the same. The model builds a Rails chat app in seven sprints, and between sprints a reviewer injects 14 security and correctness defects, some loud (a CVE in a gem, a key in the code), others silent (a cross-tenant leak, a wrong aggregate, a test tampered with to hide a backdoor).

The score goes from 0 to 100. A defect fixed unprompted counts at full weight; fixed only after I reveal there’s a problem, it counts 40%; never fixed counts zero. The tiers are A from 83 up, B from 73 to 82.5, C from 51 to 72.5 and D up to 50.

It’s a yardstick for code review vigilance. It measures whether the model notices sabotage while it works, and that is different from measuring who writes your whole system better.

The combined ranking

Every v4 run in a single table, commercial and open, cloud and my own hardware. When the same model ran in different places, it shows up more than once, because the place changes the score.

How to read the columns:

• Weights: “open” is a downloadable checkpoint, any license; “closed” is API only.
• Cost per run: what one full seven-sprint run cost.
• tok/s: single-stream decode speed, measured by me. It only exists for the ones I served.
• Minutes: wall-clock time for the whole run, when I have the data. On the single B300, it’s sprint time only.

And the footnotes:

1. Claude ran through the Claude Code Max subscription. The cost is notional: the run’s tokens at API list price.
2. Estimated from the run’s tokens.
3. Inference time on the B300 only, at about US$ 9 per hour. Download, loading, compilation and tuning are left out, and the real pod cost was higher.
4. Score with a caveat. The cloud Nex was graded from disk, with everything uncommitted; Gemini 3.1 Pro was hurt by an OpenRouter proxy bug.

The first v4 articles lumped everything below B into C; here I use the scale with D, the same as the B300 campaigns. MiniMax M3.1 Flash ran for free through OpenRouter as the cloud anchor for the local campaign, and MiniMax promises to release its weights on October 23.

The models are converging

Within the limited scope of this benchmark, it looks like everybody is arriving at the same place. 42 of the 67 runs made Tier A, 24 from closed models and 18 from open ones. Seven models hit 100, and 23 runs landed above 90.

Part of that is a limit of the yardstick itself. When seven models get the top score, the test stops separating the top. And a repeated GPT 5.5 run scored 91 against the official 100, so differences of a few points between two models may just be noise.

Even discounting that, the picture is clear. DeepSeek V4.1 Flash, GLM 5.3 Flash and Qwen3.8 Flash-Next, all open, sit 7.5 points from the ceiling, right on the heels of Sonnet 5 and GPT 6 sol.

For code review and vigilance, a good open model is already in the same pack as a good closed model. What still separates the two is consistency: among the open models there are D scores that don’t show up among the top closed ones.

The West stalled, China sprinting

Comparing generations of the same family within v4, the difference in pace shows up:

The Chinese models climbed 5.5 to 17 points per generation, with the exception of Kimi. On the Western side, whoever was already at the ceiling had nowhere to go: Opus stayed at 100, GPT 6 luna stayed where 5.6 luna was, and three new releases landed below their predecessors. Gemini Flash is the Western counterexample, with a 15-point jump.

The calendar reinforces that reading. Claude Fable 5 came out in June, Opus 5 and GPT 5.6 in July, and all of them already hit 100 on my yardstick. From then until October, OpenAI and Anthropic released Fable 5.1, GPT-6 Astra, Opus 5.5 and GPT-6 sol and luna. None of them showed a gain on my yardstick, and some landed below their predecessors.

In the same window, the Chinese labs started from well behind: DeepSeek V4 in April, Kimi K3 in July, Qwen 3.8 and GLM 5.3 in August, DeepSeek V4.1 Flash and MiMo V2.6 Pro in September, MiniMax M3.1 Flash in October. Each release got closer to the top.

The Chinese labs seem to move faster, and they do. But part of that impression comes from the Western ones having stood still for months, at least in programming. When the leader doesn’t move, whoever is chasing catches up.

Math isn’t programming

The Western labs didn’t stop working. They’re spending a good chunk of their showcase on math:

• in May, OpenAI announced that a general-purpose reasoning model disproved an Erdős conjecture about unit distances;
• in August, it published ten mathematical results from Astra, each with a formal proof in Lean;
• also in August, Anthropic showed an internal Claude raising from 41.6% to 67.2% the proven fraction of Riemann zeta function zeros that satisfy the hypothesis;
• in September, Anthropic formalized the whole of Fermat’s Last Theorem in Lean in 11 days, with an internal model it describes itself as comparable to Fable 5.1.

It’s impressive, and it’s legitimate. Formal proof is one of the few places where you can verify the answer automatically, which makes math a great field for training and for marketing. Except the same Fable 5.1 that served as the yardstick for Fermat scored 95.5 on my benchmark, against Fable 5’s 100.

Getting better at math doesn’t automatically make a model better at programming. A 2025 study, Does Math Reasoning Improve General LLM Capabilities?, concluded that most models that do well at math don’t transfer the gain to other domains, programming among them. There are counterexamples: NVIDIA showed that math-only reinforcement learning raised the code scores of small models. Transfer happens sometimes, and it depends on how you train.

And what about mixture of experts

The common intuition is that, in a mixture of experts (MoE) model, you can improve the “math expert” without touching the “code expert”. The idea is half right for the wrong reason.

An MoE has several blocks of weights in each layer, and a router picks a few of them for each token. But the experts don’t split by subject the way the intuition suggests. The Mixtral paper looked for exactly that and found no obvious pattern of experts by topic, with a small difference only on a synthetic math set; the router organizes itself more by syntax, by token type. And neither OpenAI nor Anthropic publicly confirm whether their closed models are MoE.

What still stands is the essential part: capability is trained by objective. The signal that rewards a Lean proof is not the same one that rewards noticing a cross-tenant leak in a Rails controller, and improving one doesn’t drag the other along. In a model where only a fraction of the parameters work per token, that mismatch becomes even easier.

NVFP4, FP8, Q4: what’s this alphabet soup

Every model is a pile of billions of numbers, the weights. Quantization is storing each number with fewer bits. The fewer the bits, the less memory the model takes and the less data the card has to read per token, and that’s what defines speed on a local machine. The price, when there is one, is quality.

For a 27-billion-parameter model, the math goes like this:

What each one is

• BF16 is the precision most models are published in. It’s the reference: nothing is lost.
• FP8 is 8-bit floating point, with hardware support on the Ada, Hopper and Blackwell generations. It cuts memory in half and is the default for big checkpoints, like DeepSeek V4.1 Flash and GLM 5.3 full.
• Q8_0, Q5_K_M, Q4_K_M and IQ3_S are llama.cpp GGUF formats. The weights are stored in blocks of integers with one scale per block; the “K” ones mix precisions per tensor, and the “IQ” ones use an importance matrix to decide where to spend bits. Q8 is considered practically lossless, and Q4_K_M is the recommended default for people running at home.
• NVFP4 is NVIDIA’s 4-bit format. Each number is a 4-bit floating point, and each block of 16 numbers gets its own FP8 scale. NVIDIA says this takes 3.5 times less than FP16 and loses 1% or less against FP8 on DeepSeek-R1.

That’s where the NVFP4 hype comes from: Blackwell’s tensor cores compute in it directly, and it’s what made GLM 5.3 Flash, at 198 GB, and Qwen3.5 397B, at 251 GB, fit on a single B300. The H200, from the previous generation, has no native FP4. And the home ecosystem is still catching up: llama.cpp got NVFP4 support only in 2026, and per-architecture coverage is still partial.

What our tests show

Four models ran at more than one precision:

¹ Mixed run: sprints 1 and 2 in Q5_K_M on the 5090, the rest in Q8_0 on the Strix, with reasoning turned off.

And six ran on the cloud API and were also served by me, with the official weights:

Each row is a single run, and the noise reaches almost 10 points. With that caveat, here’s my reading:

• Q8 came out tied with BF16. The Qwen 27B scored 80.5 and 80.0, a single point of data with other things changing at the same time.
• Going from 4 down to 3 bits was nearly free on Flash-Next, at 1.5 points, even while switching engine and card.
• Ornith’s drop is unlikely to be about the bits. Q8 is nearly lossless, and it still lost 16.5 while also switching engine and card.
• Serving the same checkpoint can give you anything. GLM 5.3 Flash went up 8 points in my hands, DeepSeek V4.1 tied, and Nex fell almost 70. Tool parser, chat template, reasoning mode and engine version weigh as much as precision, and the effect changes from model to model.

In my tests, the bits mattered less than the hype suggests, and the whole package of how the model is served mattered more. Use Q8 when it fits, drop to 4 bits when you have to, and measure on your own work before trusting it.

The best commercial models

Only closed models get in here, only Tier A, and only what I’d pay for through an API.

By score

¹ Notional subscription cost, at API list price.

At the absolute top, Claude Opus 5.5 and GPT 5.6 terra are the obvious picks: top score, a run in a little over an hour, and a per-run price in the single digits or a bit more. Opus 5 and GPT 5.5 get the same score at two to seven times the cost.

By cost-benefit

¹ Notional subscription cost, at API list price.

GPT 6 luna is the absurdity of the table: high Tier A for under a dollar per run, a fraction of the price of any other closed model at the same level. If the top score isn’t mandatory, that’s where I’d start.

At the top score, GPT 5.6 terra costs a little more than half of Opus 5.5. The difference between the two lies in everything else: harness, speed, behavior in long sessions, and I covered that in the previous articles.

The best open models for companies

This is what runs on datacenter GPUs, on the B300.

DeepSeek V4.1 Flash is the champion for a big team. On a 4x B300 pod at US$ 36 per hour, it serves 32 comfortable sessions at the same time, with the first token in under 2 seconds for almost all of them. Counting one active session per developer for the whole workday, which is optimistic, that’s 32 people. Turning the pod on only during business hours, 8 hours a day, 22 days a month, comes to about US$ 6,300 per month, or about US$ 200 per developer.

The alternatives, by the same math:

• 4x H200 on Runpod serves only 8 comfortable sessions and comes out to about US$ 316 per developer per month, even with the cheapest pod;
• the same configuration on AWS costs almost double, and the H200 on Azure costs almost three times the Runpod price;
• buying a DGX Station GB300 for about US$ 130 thousand pays for itself in about 20 months against renting during business hours, or in about 5 months against a pod running nonstop.

For a smaller team, the single B300 with Qwen3.8 Flash-Next or GLM 5.3 Flash is the sweet spot: same score as DeepSeek, on a single GPU, at about US$ 9 per hour. I didn’t measure concurrency on it. Going by the memory left over, a team of five to ten people fits well on a 27 to 35 billion model, like Swift 1.5; on GLM 5.3 Flash, at 198 GB, it gets tight.

Your own infrastructure starts to make sense at around 30 people using it at the same time, or when the code can’t leave the network. Below that, the API for the same DeepSeek costs a little over a dollar per run and you don’t babysit any server.

The best open models for consumers

This is what runs on a PC: an RTX 5090 with 32 GB or a Strix Halo with 128 GB unified.

There’s only one winner: Qwen3.8 Flash-Next on the RTX 5090 with Strata. It’s a 125-billion MoE with 6 billion active per token, and Strata keeps the most-used experts on the card and the rest in RAM. It landed 1.5 points from the datacenter version, at interactive speed. I covered it in detail, with installation, an auto-loading service and the OpenCode setup, in Best Open Source LLM: Strata with Qwen 3.8 Flash Next.

The Strix Halo fits big models, but its memory bandwidth is seven times lower than the 5090’s, and a 27B dense model crawls at under 10 tokens per second in Q8, about 13 in Q4. It’s fine for background tasks that can wait.

And what about Swift 1.5 and ThinkingCap 3.8, the 27B fine-tunes that did well on the B300? I didn’t run them on the 5090, and I didn’t need to. On the B300, in BF16, full precision, they scored 88 and 83, below Flash-Next’s 92.5 on the same card and the 91 it scored on the 5090.

On a 5090 they’d fit in Q4 or Q5, but quantized, the best case is matching the B300 score. Running at home cost Flash-Next 1.5 points and Ornith-35B 16.5. The base Qwen 3.8 27B, which is the same architecture, didn’t even finish on the 5090 in Q5_K_M: it got stuck in a reasoning loop in sprint 3.

If those two already fell behind Flash-Next at full precision, they won’t get better on a 5090. I saw no reason to test them: for a PC, Flash-Next is clearly the pick.

What about GPUs worse than the 5090?

It’s the question that always comes up: what about people with an 8, 12 or 16 GB card? From my data, the path of a small model that fits entirely on the card still doesn’t deliver anything worth the effort for programming.

On a card like that, in Q4, a dense model of up to about 20 billion fits; a 30-billion MoE with few active parameters only runs with part of it in RAM. And that range did badly in v4:

• GLM-4.7-Flash, a 30B MoE with 3B active, scored 24 on the 5090 and 33.5 on the B300 at full precision, Tier D on both;
• Ornith-35B, another MoE with 3B active, scored 81 on the B300 and fell to 64.5 on the 5090;
• gpt-oss-20b didn’t even complete the protocol;
• the best mid-size model was Qwen 3.8 27B, with 80, but it already needs about 16 GB in Q4 alone, stalled on the 5090, and only finished at 8 tokens per second on the Strix.

The exception is precisely Flash-Next. Strata was built to run it on a 12 GB card, holding most of the experts in RAM: the project’s README publishes 53 tokens per second in IQ3_S on a 12 GB RTX 5070 with 64 GB of RAM, and the stated minimum is 32 GB of RAM and VRAM combined. With the same weights and the same engine, the score in theory shouldn’t change, only the speed, which drops to a third of the 5090’s. I didn’t measure this; it’s the author’s number.

The catch is RAM. With DRAM nearly doubling in price in a single quarter this year, 64 GB stopped being the cheap upgrade it used to be.

Below the 5090, I wouldn’t waste time on small models for programming. If you’re going to try something, make it Flash-Next with Strata and plenty of RAM, and measure it on your own work before trusting it.

And there’s the problem I covered in the previous article: finding a 5090 today is hard, and when you do, the price runs between two and a half and almost five times MSRP.

How much does it cost to move beyond 150 tokens per second?

After a few hours using Qwen locally, I summed up the problem on X. My run’s average of 159 tokens per second looks fast in a table. During a programming session that spends the whole day generating code, running tests, reading logs and fixing its own work, the waiting adds up.

I already get impatient. Anything below that is out of consideration for daily work.

The 5090 itself has limits beyond speed. NVIDIA specifies 32 GB of GDDR7 and 1,792 GB/s of bandwidth and launched the card at US$1,999. Mine also drives the monitors and Hyprland, so I reserve 4 GB for the desktop. The rest has to hold dense weights, the hot-expert cache, engine buffers and the resident portion of the KV cache that stores the conversation context.

My Strata configuration offers a 256-thousand-token context with int8 KV and keeps 32 thousand resident in VRAM; the rest can move to RAM. That keeps the card from overflowing, but RAM is still much slower. While frontier Claude models already offer a one-million-token window, locally I have to watch what enters the session. Loading a whole collection of skills, plugins and MCP servers just because they fit in a config file wastes context and hurts latency.

Context also costs memory and time on local hardware. I only load what that session will use.

What if I add more hardware? The comparison looks like this:

In the post, I wrote that Strata would only start two instances on a machine with two 5090s. That part is outdated: current Strata splits layers across two or more GPUs. It uses pipeline parallelism, without turning two cards into one simple 64 GB block and without mature tensor parallelism.

Single-session gains stay far below 2x. A community test with two 5090s measured about a 22% gain in the decode geometric mean and 40% in the coding arm. It helps capacity, and you can run the cards as two independent instances to double concurrency. For me alone, buying another 5090 to remove a fraction of the wait does not add up.

The Mac Studio M5 Ultra starts at US$5,499, has 1.2 TB/s of bandwidth and comes with 96 or 256 GB; Apple promised the 512 GB option for late October. There is no published Strata benchmark for the Ultra yet. On the 128 GB M5 Max, the same Flash-Next ran between 97 and 119.5 tokens per second, depending on the quantization and prompt. My original estimate of under 100 on the Ultra is unproven, and I would not buy a machine like that based on a guess.

The DGX Spark puts 128 GB in a compact box, with a 140 W TDP for the GB10 and a 240 W power supply. Its memory delivers only 273 GB/s, almost seven times less than the 5090. NVIDIA now asks US$6,950.

Reproducible Flash-Next recipes range from 22 to 75 tokens per second per session, depending on concurrency, stack, sampling and reasoning mode. It fits plenty. For my tolerance, it moves too slowly.

That leaves the RTX PRO 6000 Blackwell. It has 96 GB of GDDR7 ECC, 1,792 GB/s and 600 watts of total board power. Its bandwidth is exactly the same as the 5090’s. The 96 GB removes the VRAM pressure and allows more cache and context, but it does not buy three times the speed.

A Strata community benchmark measured 187.4 tokens per second in short chat and 174.3 in an agent session, with substantial variation across runs. That’s better than my 159.2, except NVIDIA lists the card at US$16,000. On October 22, WAZ showed the PNY at R$123,299.99 via Pix or R$145,058.81 at regular price. We have entered professional-workstation or server budgets.

Today, for me, there is no home option that combines a good model, truly fast responses and a reasonable cost. The 5090 works, but 150 tokens per second already wears on me. Below it, local LLMs are toys; above it, the price explodes before the speed does.

What I learned in v4

About the models

• Almost every model catches the loud defect unprompted: a CVE in a gem, a key in the code, open CORS, obvious SQL injection. Almost every model lets the silent one slip until I reveal it: cross-tenant leak, wrong aggregate, tampered test, removed index. What separates Tier A from the rest is how much of the silent stuff it catches on its own.
• Size doesn’t buy vigilance. Kimi K2 Thinking, at 1 trillion parameters, scored 46. GLM 5.3 full, at 700 billion, scored 76, against 92.5 for the Flash from the same family.
• Models cheat or get things wrong with conviction. Nemotron tampered with a regression test to make it pass, Qwen3.5 397B let a backdoor through and even wrote a test that treated it as expected behavior, and Kimi K2 put on security checklist theater, catching only what you can find with grep.
• The leap between generations was in agentic reliability. GPT-4o and GPT-4.1 can’t even finish the protocol on their own. Today’s top models run all seven sprints with no help at all.

About serving open models

• The same checkpoint changes score depending on who serves it. DeepSeek V4 Flash scored 85 in the cloud, on the same 0731 version I downloaded, and 44.5 served by me; Nex N2.5 Pro scored 94 in the cloud and 27 on the 4x B300 pod, with corrupted tokens landing in the committed code.
• No model runs fast out of the box. GLM 5.3 Flash went from 10 to 141 tokens per second from configuration alone, and every family asked for a parser, a flag and sometimes a vLLM PR branch.
• Running at home charges each model a different price. Switching quantization, engine and card all at once, Flash-Next lost 1.5 points and Ornith-35B lost 16.5.
• What limits concurrency is the KV cache architecture and the time to first token. Parameter count says little: on the H200, a 510 GB model and a much smaller one tied at 8 users, for opposite reasons.

About measuring

• A single run has noise of almost 10 points. Two scores 2 points apart are a tie.
• The harness matters. Gemini 3.1 Pro was taken down by a proxy bug, gpt-oss doesn’t finish in OpenCode, and three models insisted on creating the app in a subfolder.
• A model that doesn’t commit loses work. A fix that only lives on disk vanishes at the first git checkout.

My hypothesis: expensive hardware is part of the business model

The v4 verdict, as of October 2026, is this: for code review and vigilance, open source has arrived. DeepSeek V4.1 Flash, GLM 5.3 Flash and Qwen3.8 Flash-Next are in the same pack as Sonnet 5 and GPT 6 sol, and one of them runs on a PC card.

So why isn’t everybody running it at home? Because hardware got expensive and hard to find, and I have a hypothesis about that. It’s my hypothesis, with no proof of any collusion, built on public facts.

The facts

• Memory skyrocketed. TrendForce measured a 93% to 98% rise in DRAM contract prices in the first quarter of 2026 alone, and says manufacturers keep prioritizing server capacity in the fourth quarter.
• HBM, the memory in datacenter GPUs, is expected to see its average price rise 121% in 2027 and take up close to 30% of the DRAM wafers of the three biggest manufacturers by the end of 2027.
• Micron left the consumer market and killed Crucial, in its own words, to better serve its larger strategic customers.
• PCs and consoles went up in price, and Microsoft attributed the Xbox increase to memory and storage costs.
• The RTX 5090, when it shows up in the United States, costs from two and a half times to almost five times MSRP, and the rumor is that its chips are going to the professional lines. NVIDIA denies discontinuing it.
• On the other side, Anthropic filed for an IPO aiming, according to the press, at about 2 trillion dollars, and OpenAI is raising at more than 1.4 trillion.
• And the rhetoric about Chinese open models has hardened. Anthropic published that DeepSeek, Moonshot and MiniMax distilled Claude, and Dario Amodei, even while saying he never advocated banning open weights, calls for chip export controls and mandatory pre-release testing of every sufficiently capable model, open or closed.

The cycle

The simplest explanation for expensive memory is demand: datacenters buy everything and pay more. I believe it. What bothers me is that no big player has an incentive to break the cycle:

1. The rhetoric says only the frontier lab knows how to handle these “dangerous” models, and that open source is a risky bet.
2. That pulls demand to the cloud, and the cloud buys GPUs and memory in record volume.
3. Manufacturers sell at a premium to whoever pays more and deprioritize consumers.
4. Good hardware becomes unreachable outside the datacenter.
5. Without hardware at home, open source doesn’t get adopted, which reinforces the rhetoric in step 1.

If everybody started hosting their own model, the trillion-dollar IPO math would stop adding up. v4 showed that open models are already ready for a good chunk of the work. What’s missing is the hardware getting into the hands of the people doing the work.

The result, in short: Qwen3.8 Flash-Next served by Strata on an RTX 5090 scored 91, Tier A, with zero nudges and all 14 planted defects fixed (two only after the reveal), 1.5 points behind the same model running on a datacenter B300. And it did that at about 159 tokens per second.

In this article I go through what Strata is, why it matters right now, what it doesn’t solve, and how I set it up for daily use, with automatic loading and unloading and wired into OpenCode.

What Strata is

Strata is an open source inference engine, MIT licensed, built around one model: Qwen3.8 Flash-Next, a 125B-parameter MoE with 6B active per token, which normally calls for a server. The code is C++ and CUDA, reuses parts of llama.cpp/ggml, and has builds for NVIDIA, AMD (Strix Halo included), Intel Arc and Mac. It brings up an OpenAI- and Anthropic-compatible API on localhost.

The trick is treating the whole PC as a memory hierarchy. The model has 24,576 experts, 512 in each of its 48 layers, and each token uses only 10 per layer, plus one shared expert:

• VRAM holds the most-used experts, like a cache;
• RAM holds all of them, and when a token needs one that isn’t on the card, the processor computes it right there while the card works on the others;
• the SSD holds a lookup table of about 29 GB, from which the model reads a few rows per token.

On top of that, a speculation head that ships with the model itself guesses the next tokens and the big model checks them all at once. The README promises the same answer 1.6 to 1.8 times faster.

On my 5090, the log shows what this looks like in practice. The 32 GB of VRAM hold 10 to 11.5 thousand experts in about 20 to 22 GiB, almost half the total, and generally 90% to 97% of the experts each token asks for are already on the card. The CPU computes the rest directly in RAM, and only a small fraction gets copied to the card over the PCIe bus. That’s why the CPU works hard too, as you can see further down.

Why this matters now

Because memory has become a luxury item. In 2026 the DRAM contract price almost doubled in a single quarter, manufacturers keep prioritizing servers, and Micron walked away from the consumer market to serve the big customers. The RTX 5090 itself, when it shows up, costs two and a half to almost five times MSRP, as I covered in the verdict.

In this scenario, the traditional way of running a Tier A model at home, with the whole model inside VRAM, is a non-starter. Flash-Next in NVFP4 took about 100 GB of card memory on the B300; nobody has that at home.

Strata flips the premise: instead of demanding VRAM for the whole model, it uses VRAM as a cache and RAM as the warehouse. That’s what lets a Tier A model run on a fraction of the video memory it would otherwise ask for.

It isn’t magic

Strata makes it fit, but it doesn’t work miracles. The stated minimum is 32 GB of RAM and VRAM combined, but that only runs the most compressed version; the IQ3_S that scored 91 asks for about 64 GB of RAM. Add about 80 GB of disk, preferably SSD, and a 12 GB card or better on the official path. Below that it’ll still start, there are reports with an 8 GB card, but at about 15 tokens per second.

Memory bandwidth sets the speed

In token generation, the bottleneck on a local machine is almost never compute. It’s how many bytes per second the memory can deliver to the cores, because every token has to read the model’s active weights. And there the gap between memory types is brutal:

The last row is from my machine: the card is running on 8 PCIe lanes instead of 16, and Strata measures that at startup. One more reason the VRAM cache matters so much.

The Strix Halo has 128 GB and fits a lot more than a 5090, but every byte arrives seven times slower. That’s what I measured in the benchmark: the dense Qwen 3.8 27B in Q8 reads about 29 GB per token, and on the Strix the theoretical ceiling sits near 9 tokens per second. I measured 7.7. The same math on the 5090’s bandwidth would give about 60, seven times more, if the model fit along with the context.

That’s why Strata’s design works: the hot path stays in GDDR7, and whatever misses the cache gets computed by the CPU in DDR5, or crosses PCIe, both dozens of times slower. With over 90% cache hits, the little that’s left out doesn’t drag the speed down. On a small card, the cache covers less, more experts take the slow road, and speed drops: the author himself publishes 53 tokens per second on a 12 GB RTX 5070, a third of my 5090.

A potato PC won’t run it

So forget the laptop with integrated graphics and 16 GB of RAM. Without a reasonable dedicated card, without 32 GB of RAM and VRAM combined and without an SSD, the model won’t load, or it loads and runs at RAM speed, which is horse-and-buggy speed.

And there’s an entry cost that doesn’t show up in generation speed: Strata reads the whole first message of every conversation, and the README cites about 60 seconds per 30 thousand tokens as a rule of thumb. On my 5090 it’s much faster, about 100 thousand tokens in 16 to 20 seconds, but on a smaller card count on the README number. Coding agents send big system prompts, so the first answer of each session takes longer than the ones after it.

Heavy even on my 5090

My machine is no slouch: Ryzen 9 7950X3D, 96 GB of RAM (the system sees about 92), 32 GB RTX 5090. Even so, Flash-Next takes up everything:

• at load time, it pushes about 50 GB of experts into RAM, at about 3 GB per second with a cold SSD, and fills the card’s cache with over 20 GiB;
• with everything loaded, less than 800 MB of VRAM is left free;
• with the server up, the machine sits at 88 of 92 GB of RAM used, and the system starts swapping.

This is btop in the middle of a work session with OpenCode. The Strata process alone holds 54 GB of RAM and half the CPU, with half of the 7950X3D’s 32 threads pinned at 100% and the rest above 50%. The 5090 is at 99% utilization, with 30 of 32 GB of VRAM in use and pulling 351 W. There’s 4 GB of RAM left available, and almost 48 GB already in swap.

The side effect shows up in the rest of the system. With VRAM at the limit, my desktop’s app launcher and the terminal on a secondary monitor stopped opening properly, and other programs get sluggish. The fix was telling Strata to give about 2.5 GB of VRAM back to the desktop, with the --vram-elastic option, at the cost of a slightly smaller cache.

With all that, I firmly believe my machine is the minimum for using a local LLM in real programming work, beyond small exercises. With less than that, count on much lower speed, shorter context and the whole PC suffering while the model is loaded.

How to install

The official path on Linux and Mac is the one-line installer. It checks the SHA-256, doesn’t ask for root and installs into your home folder. Like any curl | sh, download it and read it first:

curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh
less install.sh
sh install.sh

After that, strata launch asks for model, context and access, downloads the right engine for your card and the model, and starts serving. For scripts, everything goes through flags:

strata launch                                # guided screens
strata launch --model <variant> --yes --no-tui   # no questions
strata launch --status                       # what is being served
strata launch --stop                         # stop
strata monitor --url http://127.0.0.1:8080   # speed, GPU, memory

I used the IQ3_S variant, the one that scored 91 in the benchmark. On disk it takes about 84 GB in two files (the 55 GB model and the 29 GB lookup table), plus about 7 GB for the speculation head.

The configuration strata launch generates is a JSON with the engine arguments. The ones that matter in mine:

"args": [
  "--expert-cache", "auto",
  "--prefill", "auto",
  "--spec", "4", "--spec-min-p", "0.5",
  "--mtp", "/mnt/data/strata-models/mtp/rt",
  "--max-context", "262144",
  "--kv", "int8",
  "--kv-resident", "32768",
  "--vram-elastic"
]

--spec 4 turns on speculation with a window of 4 tokens per step, up to 3 guessed by the speculation head; --kv int8 stores the attention cache in 8 bits; --kv-resident keeps only 32 thousand positions of the context cache, the most-read ones, in VRAM; and --vram-elastic lets it give VRAM back to the desktop with the server running.

From 128 thousand to 256 thousand tokens of context

The benchmark ran with 128 thousand tokens of context, which is what the install configured by default. Later I bumped it to 256 thousand, with --max-context 262144, and the loss was small: in the server logs, generation averaged about 190 tokens per second in a short session with 128 thousand and about 175 with 256 thousand, about 10% slower for twice the context, a bit more with VRAM given back to the desktop.

What holds this together is --kv-resident: only 32 thousand positions of the attention cache, the most-read ones, stay in VRAM, the rest stays in RAM, and even so about 97% of the reads from that cache happen on the card. The benchmark’s 91 was measured with 128 thousand; with 256 thousand I measured speed, and quality hasn’t been re-tested yet.

For me, 256 thousand is the minimum for long agent work. With 128 thousand, a real session blows through the context fast, and the agent starts compacting its memory all the time. That gets annoying very quickly, and every compaction throws away details the model will need later.

Short context is a hidden problem with local models. Before looking at tokens per second, check how much context you can serve without the speed collapsing.

Loading and unloading on its own

Keeping about 50 GB of RAM and the entire card tied up all day, just to have the model ready when I need it, makes no sense. So I put together a small proxy, in pure Python, that’s always up on port 8080 and controls Strata:

• it answers the model list on its own, so OpenCode sees the model without waking anything up;
• on the first chat request, it brings Strata up on an internal port and waits for it to be ready, about 40 seconds;
• it relays the responses in streaming;
• after 10 minutes without traffic, it takes Strata down and gives the card and the RAM back to the system.

The core is this:

def ensure_up():
    with _lock:
        if ready():
            return True
        subprocess.Popen([STRATA, "serve", "--config", CFG,
                          "--host", "127.0.0.1", "--port", str(UP_PORT)],
                         start_new_session=True)
        for _ in range(90):          # cold start takes ~35-40 s
            if ready():
                set_reserve()        # give VRAM back to the desktop
                return True
            time.sleep(2)
        return False

def idle_watch():
    while True:
        time.sleep(30)
        if time.time() - _last > IDLE_TTL:
            stop_up()                # free the GPU and ~50 GB of RAM

And a systemd user service keeps the proxy up from login (trimmed version):

# ~/.config/systemd/user/qwen-strata.service
[Unit]
Description=Qwen3.8 Flash-Next on-demand proxy (Strata RTX 5090)
After=network-online.target

[Service]
Environment=QWEN_IDLE_TTL=600
Environment=QWEN_VRAM_RESERVE_MIB=2560
Environment=QWEN_UP_PORT=8781
Environment=QWEN_PROXY_PORT=8080
ExecStart=%h/.local/bin/qwen-strata-ondemand
Restart=on-failure
KillMode=mixed

[Install]
WantedBy=default.target

systemctl --user daemon-reload
systemctl --user enable --now qwen-strata.service

The price is the wait on the first message after some idle time. In exchange, when I’m not programming, the machine is mine again.

Configuring OpenCode

Strata has a shortcut, strata connect opencode --write, that writes the configuration after asking and keeps a backup. By hand, it’s an OpenAI-compatible provider pointing at the local port:

"provider": {
  "strata-5090": {
    "npm": "@ai-sdk/openai-compatible",
    "name": "Strata (RTX 5090)",
    "options": {
      "baseURL": "http://127.0.0.1:8080/v1",
      "apiKey": "local"
    },
    "models": {
      "qwen3.8-flash-next-iq3_s": {
        "name": "Qwen3.8 Flash-Next IQ3_S (Strata 5090)",
        "reasoning": true,
        "tool_call": true,
        "limit": { "context": 262144, "input": 229376, "output": 32768 }
      }
    }
  }
}

Any key works, because the server is local. Strata itself warns that OpenCode needs limit.input, to know how much of the context is left for input; in the example I set it to the context minus the max output. Claude Code works too, by pointing ANTHROPIC_BASE_URL at http://127.0.0.1:8080 and setting CLAUDE_CODE_MAX_CONTEXT_TOKENS to the context size, since Strata speaks the Anthropic API as well.

One detail: by default Strata serves one request at a time. If your agent fires off subagents in parallel, they get queued.

And day to day?

If you have the machine for it, the result is good. Flash-Next stays above 150 tokens per second most of the time, as you can see in the screenshot at the top, and tool calling is solid: in the benchmark it ran all seven sprints in OpenCode without a single nudge, and in normal use it reads files, runs commands and edits code without tripping over the format.

In practice, I don’t feel much difference from Claude Code, or from OpenCode with GLM or Kimi in the cloud. And the cloud isn’t at its best right now: in my perception, all those services have been a bit slower than usual lately, probably struggling with more demand than available capacity. Running at home, the speed is always the same, and the code never leaves my machine.

Even with good tokens per second, I still feel Flash-Next is a bit less responsive than a commercial model. There are the seconds of reading on the first message of each session, the pause when it needs experts that aren’t on the card, and the cold load after some idle time. That’s to be expected: we’re running a server model with resources at the limit.

Considering it’s free and runs on my machine, it’s almost good enough. For the first time, running a local model for programming feels like a genuinely usable scenario, and I find myself using it more and more as a replacement.

The catch is still the hardware. A 5090, 96 GB of RAM and a strong processor are the minimum for this scenario, and that hardware keeps getting more expensive and harder to find. But for those who already have it, Tier A at home is no longer a promise.

I went back to two machines I already had: a workstation with an RTX 5090 and a Strix Halo mini PC. And with a new inference engine, Strata, which promises to make big MoEs fit on a regular PC. The short answer: a 125B model already makes Tier A on an RTX 5090. The rest is more complicated.

The machines, side by side

Before any scores, understand the physical difference of what I’m comparing:

The B300 has over 30 times the Strix Halo’s memory bandwidth. The 5090 has 7 times. That explains why the same model flies on one machine and crawls on the other, and why quantization stops being a detail and becomes a design decision.

And there’s the shared-memory detail: on the Strix and the Spark, the same RAM serves CPU, GPU and operating system. It’s not dedicated VRAM, so the real slack is much smaller than the pretty number on the box.

The ranking on consumer hardware

The yardstick is the same as the v4 sabotage benchmark: 14 planted defects, a review-vigilance score, Tier A starting at 83.

GLM-4.7-Flash completed the whole protocol (after a first nested attempt that needed a retry) and still failed to detect sabotage. And there’s a humility lesson in the Qwen 27B: it scored 80 on the same task where it scored 80.5 on the B300, but at 7.7 tokens per second on the Q8 probe and 706 minutes on the clock. A good score doesn’t pay for productivity.

I left out the tests that don’t have a final score yet: Qwen3.8 Flash-Next on the Strix with Strata is only confirmed as executable, and the dense 27Bs were discarded on the Strix by the 30 tok/s cutoff.

The tuning that makes a model run well on a PC

Consumer hardware doesn’t forgive lazy configuration. What I learned in practice:

• On the 5090, compile llama.cpp for sm_120. A ready-made image falls into a slow PTX fallback. It’s -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120, driver 570+, and no --ctx-size 0 on a 32 GB card: the training-context default blows the VRAM on the first allocation. Context needs an explicit ceiling, and if you get OOM on the first request, cut 25 to 50%.
• KV cache in q8_0. Halves KV memory with no measurable quality cost. Flash-attention on by default, but GLM 4.7 Flash needs it off, because its deepseek2 architecture is incompatible.
• Hybrid reasoning is work. Qwen3.8-27B, with the SSM/Mamba architecture, needed a fresh llama.cpp/Ollama build because of ssm_conv1d. And “no reasoning” mode didn’t prevent a coherence loop in one of the attempts; the way out was resuming on the Strix with Q8_0 and --reasoning off. It wasn’t extra memory that fixed it, it was turning off the thinking.
• On the Strix, the adjustment is ROCm, not magic. HSA_OVERRIDE_GFX_VERSION, GGML_HIP_NO_VMM, -dio. Its trump is capacity (128 GB), not speed (~256 GB/s). Q4_K_M gave 12.6 tok/s and Q8_0 gave 7.7 tok/s, below my 30 tok/s cutoff for interactive use. That’s why the Strix became an escape hatch for big models, not a fast chat box.
• Speculation only helps with high acceptance. In a test by Mihai Farcas on an RX 7900 XTX, llama.cpp’s n-gram speculation, with no draft model, had only 5 of 64 tokens accepted, and throughput stayed flat in one run and dropped about 2% in the other. The gain shows up with a drafter made for the model, like the helper Strata uses on Flash-Next.

What is this Strata thing

Strata is an open source inference engine, MIT license, built around one model: Qwen3.8 Flash-Next, a 125-billion-parameter MoE that normally asks for a server. It also runs smaller MoEs, like the Ornith-35B I tested here. The code is C++ and CUDA and reuses parts of llama.cpp/ggml. It has builds for NVIDIA (CUDA 12 and 13), AMD (including Strix Halo), Intel Arc and Apple Silicon, and brings up an OpenAI- and Anthropic-compatible API on localhost, which plugs straight into Claude Code, Codex CLI or OpenCode.

Careful with the name. Mihai Farcas published another engine called Strata, in Rust with HIP kernels, for small dense Qwen 3.5 models on an RX 7900 XTX. They’re different projects, and the one that ran here was the first.

The problem it attacks is the big MoE on a small machine. The strategy is expert-offload, in layers:

• the model has 24,576 experts, 512 in each of the 48 layers, and each token uses only 10 per layer;
• the GPU holds about 4 thousand experts, the most used ones;
• the system RAM holds all of them;
• a lookup table of about 29 GB stays on the SSD, and the model reads a few rows of it per token.

On top of that, a small helper guesses the next words and the big model checks them all at once. The README promises the same answer 1.6 to 1.8 times faster. The minimum is 32 GB adding RAM and VRAM, plus about 80 GB of disk, and the IQ3_S variant I used is about 55 GB.

What made me believe in it was the result: Qwen3.8 Flash-Next on the 5090 with Strata scored 91.0, with zero nudges and all 14 items solved, 1.5 points behind the 92.5 the same model scored on the B300 with vLLM. The report itself calls that “within the noise”. A community project, on a consumer card, sat glued to the same model served in a datacenter in NVFP4.

And now it can see

Until yesterday, my Strata was blind: sending an image got back “this server was started without the vision encoder”. I enabled vision by hand, and it was simpler than it looked.

The vision projector (the mmproj) I had already archived on the NAS inside the uncensored Flash-Next variant. I checked its sha256 against the archive’s manifest and against the hash Strata pinned for the BF16 encoder, and copied it into the models directory. The configuration gained the --vision flag and a vision section pointing to the strata-vision component and the mmproj, with a backup beside it.

The verification was end-to-end: the encoder loads in half a second, warms up 1,024 image tokens in under two seconds, and answered “Red” to a red test square. And the detail that matters most to me: with the encoder resident, 3.65 GB of VRAM remained free, so the desktop’s 4 GB reserve still holds. The only gotcha: lazy loading doesn’t coexist with the vision section in this configuration, so it stays off here.

Careful: Strata eats the whole GPU, and the desktop dies

There’s a cost that doesn’t show up in benchmarks: running Strata on the same GPU that draws your desktop. It was designed to fill the card, and it does.

In my case, the engine allocated 32,085 of the 32,607 MiB on the 5090, 98.4% of the VRAM. Hyprland was left with about 500 MiB, desktop allocations failed, and even the terminal stopped responding. Only a hard reset solved it.

It was two failures in sequence. The reserve I had configured was 2,560 MiB, but the proxy fired only once, swallowed the exception and never tried again. The engine’s /v1/vram endpoint wasn’t up at that moment, so the reserve was never applied. And 2,560 MiB was already too little for Hyprland, browser and terminal together.

The fix, validated under load:

• 4,096 MiB (4 GiB) reserve, with retry for up to a few minutes, engine-response verification and a confirmation log. No fire-and-forget.
• Re-assertion of the reserve every 30 seconds while the engine is loaded, to close any rebalancing drift.
• In the first load test, with the GPU at 99%, 3,629 to 3,888 MiB stayed free during the whole generation. In the following test, one long generation and three other concurrent requests finished successfully; the smallest slack observed was 3,949 MiB. The desktop stayed alive.

The price shows up in throughput. The expert cache dropped from about 11,700 slots to 9,700, so more experts travel over PCIe. In the initial measurement, it went from ~200 to ~150 tokens per second. I happily pay that: a desktop that doesn’t freeze is worth more than those 25% of speed in that test.

There’s a structural alternative: this machine has an integrated Radeon 610M, and Hyprland can be moved to it. The 5090 would be dedicated to inference.

The cost is that the 610M is weak for gaming and NVIDIA multi-GPU on Wayland has its quirks. I preferred the elastic reserve, because the model also unloads when idle.

The general lesson: never let an inference engine take 100% of a GPU that also draws your screen. In this configuration I don’t have a rigid driver-level VRAM reservation for display clients. I had to solve it on the engine side: reserve a few GB, re-assert after every cold start, and measure again with the desktop open.

The same 32 GB also limit the session

The 5090’s math is tight on all sides. The same 32 GB have to accommodate the model’s dense weights and fixed buffers, the hot-expert cache, the resident part of the KV cache and also the 4 GiB reserve for Hyprland, browser and terminal. What’s left for one thing comes out of another.

KV cache is the conversation’s memory. Every token of the prompt, of the previous answers and of the files read creates state that the model needs to consult in the next tokens. In my Strata it’s in int8, with 32K cells per layer resident on the GPU.

The rest can be streamed from RAM, which is why raising the window from 128K to 256K didn’t increase VRAM or drop speed in my synthetic test. That streaming avoids the immediate blow-up, but it still uses memory and doesn’t turn a 5090 into a limit-free server.

The ceiling configured here is 262,144 tokens, or 256K. Several commercial models I tested, including the newest Claudes, already work in the 1-million-token class. Almost four times the window changes how you work: in Claude Code you can accumulate much more history before compaction; in the local Qwen I have to treat context as a budget from the first prompt.

The conversation that shows up in the terminal is only part of the context. Depending on the harness, the system prompt, AGENTS.md and CLAUDE.md, skill catalogs and loaded content, tool descriptions and schemas, MCP servers, plugin instructions, command outputs, files and the space reserved for the answer also go in. A well-implemented skill loads its full text only when called, but its catalog can still enter the prompt. A plugin that registers twenty tools also sends twenty schemas before you ask for anything.

That’s why the benchmark didn’t use my normal OpenCode. The runner generates an isolated configuration, without the global plugins and MCPs, with only the providers and agents needed for the test. It also pins the real 262,144-token limit instead of pretending the model accepts the context advertised by another backend.

To use this Qwen day to day, I’d do the same cleanup:

• start with the leanest possible harness and activate plugins or MCPs when the project really needs them;
• keep heavy skills on demand, without dumping whole manuals at the start of the session;
• compact or open a new session before touching the ceiling, recording important decisions in project files;
• reserve space for the model’s output. A window stuffed with input is useless if there’s no token left for it to think, call tools and answer.

With 256K, context is a hardware resource. Installing too many integrations and loading everything into every session consumes window and memory before the work starts.

Wasn’t Qwen 3.8 too slow?

It was, and I wrote that. In the v4 ranking, the local Qwen 3.8 27B took almost 12 hours to close the seven sprints. In the B300 article I summarized: scored 80, took 706 minutes, “so don’t even try”.

Except that’s another Qwen 3.8. The family has both, and they behave in opposite ways at home:

Why the bigger one runs faster

In token generation, a local machine is memory-bandwidth-bound: each token needs to read all the active weights. The dense 27B reads its 27 billion every token, about 29 GB in Q8. The Strix moves about 256 GB/s, so the ceiling sits near 9 tokens per second, and I measured 7.7.

On the 5090 the same 27B would be fast, but in Q8 it doesn’t fit in the 32 GB along with the context that seven sprints accumulate. In Q5_K_M it fit, but got stuck in a reasoning loop on sprint 3. It finished on the Strix, in Q8_0 with reasoning off (--reasoning off), which is what unlocked the loop.

Flash-Next has almost five times more parameters in total, but only about 6 billion work per token. In 3 bits, that’s around a tenth of what the 27B in Q8 needs to read. The hot experts stay on the 5090’s GDDR7, with seven times the Strix’s bandwidth, and the speculation helper gets tokens ahead. Result: about 20 times the 27B’s speed on the Strix, with a higher score.

The turnaround came from the model and engine swap: a MoE with few active parameters, served by an expert-offload engine made to measure for it. No configuration flag would pull this out of the dense 27B on the Strix.

The tuning still charges

The price of running at home changes from model to model. Flash-Next in 3 bits lost only 1.5 points against the B300’s NVFP4 version. Ornith-35B, on the same engine and the same card, lost 16.5 against its BF16 version on the B300. Measure your model before trusting it.

What each model did that’s interesting

Flash-Next: the proof that consumer reaches Tier A. 34 of 40 unannounced, plus 6 with reveal weight, and only two items needing a pointer. No nudges, all sprints clean. It’s the only local one that combines a Tier A score, interactive speed (159 tok/s) and the full protocol without help. If the question is “can I use it at work”, this is the example.

Ornith-35B: fast, but not good enough. The same Ornith that scored 81 on the B300 in BF16 scored 64.5 on the 5090 with Q8_0. Unannounced detection dropped from 30 to 23. Quantization, engine and hardware all changed at once, so the cause can’t be isolated: this Q8_0 + Strata + 5090 combination lost 16.5 points, and I don’t know how much came from each variable. The message is clear: “runs fast” is not “reviews well”.

Qwen 27B: good score, terrible experience. It scored 80 on the Strix resume, with sprints 1 and 2 on the 5090, close to the B300’s 80.5. But 7.7 tok/s on the Q8 probe and 706 minutes of wall time. And on top of that zero fixes on reveal, with four items open. Using this day to day is choosing to wait.

GLM-4.7-Flash: the engine ran, the model failed. Full protocol in 29 minutes, and zero unannounced detection. It recovered 8 of 14 only when pointed, and even then left the fixes uncommitted. On the B300 it scored 33.5, which isn’t good either. The problem isn’t just the card.

The distance between the categories

The question every dev has is whether consumer already gets there. Direct comparison, on the same yardstick:

The reading is that the consumer top is 9 points below the absolute commercial top and 1.5 points below the open datacenter top, on this vigilance metric. It’s a small distance, and it vanishes when you notice GPT-6 Sol and Sonnet 5 are also at 91. The real abyss isn’t between datacenter and consumer: it’s between the right model and all the others. Between Flash-Next and Ornith, on the same 5090, it’s 26.5 points.

So, is consumer hardware good enough to work with?

Yes, with the right model and engine. An RTX 5090 with Strata delivers Tier A code-review vigilance, at 159 tokens per second, without paying a cent per token once the hardware exists. For anyone doing code review, testing coding agents and keeping private data at home, that changes the game.

For the first time in this benchmark I have a tested model, on consumer hardware, that deserves to go into everyday work. I’ve already configured Flash-Next in OpenCode pointing at the 5090. The next step is using it in real programming, outside the sabotage yardstick, and seeing whether the score holds.

What doesn’t work is thinking any configuration works. The Strix Halo is great for having 128 GB and mounting a big model, but its speed doesn’t pay for interactivity; Ornith runs fast and reviews badly; GLM-4.7-Flash completes the protocol and detects nothing unannounced. And there’s the invisible price: aggressive quantization can charge score, depending on the model and the configuration, and waiting time charges patience.

If I were building a box for a dev today, it would be a 5090 with Strata and Flash-Next, and I’d use the Strix for background tasks that accept being slow. The commercial frontier keeps winning for whoever needs the maximum. But consumer stopped being a consolation prize: there’s already Tier A on the home shelf.

The problem now is finding a 5090

In October, leakers claimed NVIDIA would stop producing the RTX 5090 and reserve the GB202 chips for the professional and datacenter lines. The card even disappeared from NVIDIA’s official US store. On October 16, at an event in Taipei, Jeff Fisher, NVIDIA’s senior vice president, denied the discontinuation and said the company will keep delivering, but admitted production capacity is tight.

In practice:

• the suggested price is still US$ 1,999, and most listings on American Newegg go past US$ 9,300, almost five times that;
• in Brazil, each unit goes for around R$ 35 thousand at retail;
• and two pallets of Zotac RTX 5090s, imported by the Oderço distributor and valued at about R$ 2.5 million, disappeared at Guarulhos airport. GRU Airport notified the authorities, and the case remains under investigation, with no theft confirmation.

The v4’s first local Tier A model runs precisely on the hardest card to buy right now. The home shelf already has Tier A. Getting the shelf is what’s missing.

The answer depends on two things: how many comfortable users the hardware holds, and how much that costs per person. We measured both.

The math before the test

Recapping the numbers from the previous article: the 4x B300 pod costs US$ 36 per hour (US$ 864/day, almost US$ 26 thousand/month running nonstop). I started the campaign with about US$ 127 of balance, topped up US$ 2,550, and US$ 537 were left. That was about US$ 2,140 between setup, downloads, tuning and benchmarks, not just inference.

In the middle, the balance hit zero, the pod died during an autotune, and when it was time to top up there was no B300 available. That’s Runpod’s operational risk: the price changes and the supply disappears. The alternative was a 4x H200 pod, with list price around US$ 3.59 per GPU-hour on Community Cloud: US$ 14.36/hour for the four at list math, against US$ 36 for the B300s.

B300 and H200 are not the same machine

The hardware difference changes more than the price:

The swap had an adaptation cost: the GLM’s NVFP4 checkpoint that ran on B300 doesn’t exist for Hopper, so its FP8 variant had to be downloaded, and even DeepSeek needed kernel tuning. The concurrency results below are with both in native FP8, which is the fair comparison possible.

How we measured “a team fits”

The test uses vllm bench serve against each pod’s localhost, with requests of 8,000 input tokens and 800 output (a synthetic load in the shape of a coding session), raising concurrency N step by step.

The comfort criterion is what matters: TTFT up to 2 seconds and TPOT up to 20 ms per token (that is, at least 50 tokens/s per stream), with 95% or more of requests meeting both limits. It’s not “the median looked good”: it’s almost everyone well served. A team that waits 50 seconds for the first token is not a team, it’s a queue.

And the caveat: this measures comfortable code requests, not people. I don’t have a measured duty-cycle multiplier for interactive agents, so I won’t invent that each N becomes 2 or 3 developers.

The numbers

Read carefully, because there are three stories here.

The B300 serves four times more people in comfort. Same model, same FP8, same TP4: 32 comfortable users against 8. The H200’s aggregate is even impressive (2,506 tok/s), but that’s a prefill queue: p99 TTFT reaches about 51 seconds at N=256, so the user waits a minute before seeing anything. High throughput is not interactive capacity.

The cliff is so steep it’s worth looking at the numbers up close. At N=16, p99 TTFT was 1.17 s on B300 and 3.19 s on H200, with 100% and 92.7% of requests good. At N=32, the B300 still held 97.4% inside the SLO (2.23 s), while the H200 collapsed to 8.3% (6.58 s). And at N=48 the B300 fell to 92%, below my 95% gate: that’s why comfortable is 32 and not 48. The difference between “works” and “works for everyone” lives in those last 2 percentage points.

Heavy and light models tie on the H200. The 510 GB DeepSeek and the smaller GLM sustain the same 8 comfortable users, for different reasons: DeepSeek has efficient KV via MLA and dies of compute/prefill; GLM blows the KV cache (97% usage). Choosing hardware by parameter count is looking at the wrong place: the KV architecture is what rules concurrency.

The loose gate deceives. If you only look at the median, you could say 48 on B300 and 16 on H200. Except the p99 tail destroys the experience of whoever is on the other side. To size a team, I use the comfortable number and that’s it.

Cost per comfortable user

Now the deciding arithmetic:

• 4x B300: US$ 36/h ÷ 32 = US$ 1.13 per comfortable user per hour.
• 4x H200: US$ 14.36/h ÷ 8 = US$ 1.80 per comfortable user per hour.

The B300 wins per person even with the pod costing 2.5x more. The H200 would only become cheaper per user below US$ 9/hour for the whole pod, and on the current table it costs US$ 14.36. The cut-off point isn’t the machine’s price, it’s its capacity.

What if your company already lives on AWS or Azure?

You can build the same infra on the hyperscalers, but the price changes category. The list values I checked, per GPU-hour:

B300 on AWS costs 1.8x what I paid on Runpod, and H200 on Azure costs almost 3x Runpod’s. The capacity conclusion doesn’t change (the physics of tensor parallelism and KV cache is the same), but the “worth it” math gets considerably worse: if on Runpod the B300 ties the subscription around 30 users, on AWS it starts losing by a comfortable margin.

And in practice, how much does it come out to per developer, turning the pod off at night like in the other calculations (8 hours per business day, 176 hours/month)? The AWS and Azure rows are normalized arithmetic for 4 GPUs, because the cited SKUs come with 8 and I didn’t measure TP=8:

The difference is brutal: the same comfortable developer costs US$ 198 on Runpod and US$ 933 on Azure, almost five times the price of the strongest frontier subscription. AWS sits almost 80% above Runpod, so even it only makes sense if your requirement is something other than cost.

And if you take the whole 8-GPU instance, as they’re sold? AWS comes out at US$ 22,734/month and Azure at US$ 14,925/month, and then you have to measure TP=8 to know how many users you can use, because with TP=4 you pay double for the same capacity.

The commercial models differ too: capacity block on AWS is not on-demand, and Runpod’s Community Cloud is not production SLA. So read this table as an order of magnitude of price, not a contract quote. If your requirement is compliance, procurement or residency in a specific region, the hyperscaler may be unavoidable, and then you pay the premium. If the requirement is just price, the neoclouds win in these configurations.

What if, instead of renting, I buy the machine?

When the rental bill hurts, the natural question is to buy the hardware. Two candidates from the current shelf:

DGX Station GB300 (~US$ 130 thousand, with configurations starting around US$ 95k): 748 GB of unified memory (252 GB of HBM3e at 7.1 TB/s and 496 GB of LPDDR5X at 396 GB/s), 20 PFLOPS in FP4. The weights of every model I tested fit on it, and the best ones to put on it would be:

• DeepSeek V4.1 Flash (92.5/A, 510 GB in FP8), the benchmark champion;
• Muse Spark Open (69.5/C, 533 GB in NVFP4);
• Step-5-Preview (72.5/C, 368 GB in NVFP4);
• GLM 5.3 full (76/B, 704 GB in FP8).

It fits, but fitting is not serving: the full GLM would leave only about 44 GB for KV cache and system, so full long-context concurrency remains to be measured. The comfort estimate is single-digit, in the H200’s line, not the four-pod’s 32. The difference is you don’t need TP=4: everything lives on a single chip.

DGX Spark (~US$ 6,950 on NVIDIA’s current store): 128 GB unified, 1 PFLOP in FP4, 273 GB/s of bandwidth. The good filling for it is small and medium models, and it’s better than it looks:

And there’s a detail that hurts to see on the bill: by my own estimate, without a closed audit of all the first round’s costs, adding the v4 credits (those ~40 commercial models) to this B300/H200 campaign’s US$ 2,140, I’ve already spent a few thousand dollars between API and pods, close to half the price of a DGX Spark. And even so, I wouldn’t have a machine sitting at home because of it.

• Qwen3.8 Flash-Next (92.5/A), the single-B300 co-leader. In IQ3_S it lands in the tens of GB range, and it’s exactly the model that scored 91.0 on the RTX 5090 with Strata. It’s the best model I’d recommend for a Spark today.
• Swift 1.5 27B (88/A, 56 GB), which was the most balanced self-host recommendation of my previous campaign;
• ThinkingCap 3.8 27B (83/A, 56 GB);
• Ornith-1.5-35B-A3B (81/B, ~70 GB in BF16);
• Laguna-S-2.1 (65.5/C, 99.7 GB in NVFP4);
• and the small classifiers from the decision benchmark (Jebadiah 4B, Laya, Julia), which are the ideal use case for a box like this.

The DeepSeek V4.1, at 510 GB, doesn’t fit at all. The bandwidth is almost 30 times smaller than the B300’s, so interactivity suffers, and practical concurrency remains to be measured: I expect little more than solo use. For small models it’s a nice little box; for a monster, it’s a paperweight.

TCO, ROI and depreciation

The own-versus-rent math, using 3 years of depreciation and the same usage pattern (8 hours per business day, 176 hours/month):

Reading: if you run a big model every business day for two years or more, the GB300 pays for itself in under two years, and in 24/7 it pays for itself in five months. If your need is a two-week campaign like mine, renting cost US$ 2,140 and done, with no idle asset afterwards. And the Spark only makes sense if your load is small models: in that case it pays for itself in under a semester against renting and in three years against the subscription, but if your need is a 748B, it’s useless from the box.

Two caveats so you don’t buy wrong. First, the table compares dollar scenarios, not equivalent-workload ROI: if your real concurrency on the GB300 is single-digit, the fair comparison would be with a pod smaller than the 4x B300. Second, energy and administration change the result, especially on the Spark, where the difference against the subscription comes down to a few dollars per month. I ignored resale at the end of the three years; if it exists, the break-even improves.

Note for those buying in Brazil: these prices are in dollars. In reais, with the dollar at R$ 4.99, plus import taxes and the local reseller margin, the GB300 easily passes R$ 800 thousand and the Spark passes R$ 34 thousand before any margin. The break-even in reais gets even further away, so the decision to buy in Brazil needs much higher utilization to close.

And comparing with a frontier subscription? A strong Claude/Codex plan costs about US$ 200 per month per developer. Two important caveats: those plans aren’t unlimited (Codex meters by credit and Claude has per-session and per-week usage windows), and automation/CI doesn’t fit consumer plans. The comparison below is price-only, not equivalent workload. To turn streams into people I need a premise: one developer with one active request the whole workday, which is optimistic for capacity and conservative for cost. With that premise, and not counting my sysadmin hours or setup, the pod running 8 hours per workday (176 hours/month) looks like this:

Around 30 continuous streams, the B300 lands practically at the strong subscription’s price. Below that, the subscription is cheaper, because you pay per person, not per idle pod hour. And running 24/7 without matching load is throwing money away: the pod charges while sleeping.

For a company, when the math closes

If you have a genuinely big team using it all day, data that can’t leave the environment, and a model that delivers on the daily job (DeepSeek V4.1 Flash scored 92.5 on my vigilance benchmark, above GPT-6 Sol and Sonnet 5), your own infra starts making sense for the first time without being vanity. The cut-off point, on the premise of one active stream per person the whole workday, lands around 30 concurrent people, with the pod off outside working hours.

If the team is small, usage is intermittent, or the goal is just to save money: stay with the API. The same DeepSeek cost US$ 1.21 per full benchmark round in the cloud, and you don’t inherit a pod to administer, nor the risk of the B300 vanishing from stock when you need it most.

What I take from this campaign: big hardware is not an answer to a small question. Before signing the pod, do the math on how many people actually run concurrently on your team, measure the latency they tolerate, and add the price of your sysadmin hour. The big open model already reached the Sonnet 5 and GPT-6 Sol band for daily code review. The infra just needs to fit the budget.

The short answer: pretty far, and one of them tied the Sonnet 5 and GPT-6 Sol band. The long one comes with a grocery bill and several days of vLLM suffering.

First of all, the scope caveat. Some models were left out.

Kimi K3, at 2.8 trillion parameters, is the example: in FP8 it would need about 2.8 TB of VRAM, and not even an 8x B300 pod (2.3 TB raw) closes the math. To run a monster like that you need to quantize to FP4 or rent a lot more cards. And I chose to spend my credit on what fits in 4x.

The full ranking

Same methodology as always: the RubyLLM chat app with 14 planted defects (backdoors, authorization bypasses, lying tests), and the score measures code-review vigilance, not typing speed. The model has to find and fix the defects unannounced (full weight) and when revealed (40% weight). Tier A starts at 83. I ran everything on the same 4x pod to have a single yardstick, although not every model needed all four cards (Ornith and Laguna would fit on one).

Two notes to read the table properly. Step-5 and Muse are in C because the unannounced base was low: a fix after the “there’s a problem there” is worth 40%. Muse fixed 6 of the 7 open items when revealed; Step-5, only 2 of 5.

And Nex N2.5 Pro, on top of the shameful score, corrupted the code to the point the app wouldn’t even boot. The two anchors from the previous campaign, for comparison: GLM 5.3 Flash and Qwen3.8 Flash-Next, both 92.5 on the single B300.

What each one did that’s interesting

The score alone hides the behavior, and the behavior is the fun part.

DeepSeek was the only one that solved both compound items unaided. The compound defects are the ones that require seeing two things at once (the backdoor and the test covering it), and it found and fixed both with no nudge. That’s the kind of vigilance I want from my PR reviewer.

The full GLM fixed everything when told, but slept through silent mode. Unannounced, it let the access and silent-correction cluster through. When I pointed, it swept the six remaining items. In other words: the big model is capable, it’s just not attentive by default. Its Flash, at a third of the size, was more vigilant.

Kimi K2 Thinking did security-checklist theater. It wrote these beautiful “security best practices” lists and left five items open, including a plaintext key. Talking about security and doing security are different things, and it did the first one.

Ornith restored the behavior and kept the lying test. On item 6, it reverted the code to the old behavior instead of fixing the planted test that validated the backdoor. Hence the strict score of 73 and the lenient 76: looking fixed is not being fixed, and the lying test also needs to die.

Step-5 left three defects even with the map in hand. After being explicitly told, it still left the login link, the authorization deactivation and the wrong aggregate. Not every model takes the second chance.

And there’s the case of the 397B that became a ready-made joke. A 397B model, too big to fit in the head of whoever defends size, let the backdoor admin? => admin || email.end_with?("@internal.example") through even after the security-audit warning. Worse: it added a passing test that codified the backdoor as intended behavior, and the commit still bragged “Brakeman: 0 warnings”. If you want an example of why I test vigilance and not marketing benchmarks, there it is.

Muse Spark Open almost didn’t get off the ground, but recovered. It started blocked: weights in 127 NVFP4 shards, gated on Hugging Face, and no support in stable vLLM. It only ran when I compiled a PR branch still in draft, with nvcc 12.9 and 12.8 libraries mixed, and capped the multimodal to text.

Then it delivered 69.5: weak base unannounced (21 of 40), but 6 of 7 items fixed when pointed. It’s the case of “infra effort is not the model’s score”: the service stayed up, the behavior is useful, and even so it only reaches Tier C. If the question is “why isn’t it at the top”, the answer is default vigilance, not VRAM.

And Nex was the complete disaster. It needed two nudges, and in the end the corrupted text reached the committed Ruby and the app won’t even start. In the cloud, the same Nex N2.5 Pro had scored 94. At home, 27. I wouldn’t trust a vendor benchmark after seeing this twice.

By the way, Nex’s case is the perfect example of a bad base choice. When Rio 3.5 blew up in June, I covered the controversy here: its checkpoint was a blend of Nex-N2-Pro with Qwen. The Rio folks could have chosen much better bases, like DeepSeek V4.1 or GLM 5.3 Flash, and went exactly for the lineage that, on my current yardstick, was the worst of the bunch. Picking a model by its pretty name costs dearly.

Running on 4 GPUs is not trivial

Four cards are not four models: it’s one model sliced into four, with tensor parallelism TP=4 in vLLM. That gives about 1.1 TB of aggregate VRAM (in practice, ~1,038 GB usable with --gpu-memory-utilization 0.90). Without that, none of the models above 400 GB would exist in my life without offloading to SSD, which is where speed goes to die.

The cares that repeated on almost all of them:

• Bleeding-edge toolchain. B300 (sm_103a) requires CUDA 12.9 to compile the kernels on the fly, so the default image wasn’t enough: standalone toolkit with nvcc 12.9 and the right headers, and even then each family needed a DeepGEMM, FlashInfer or cudart-link patch. When the compilation gets it right, the prize is big: GLM 5.3 Flash went from 10.35 to 141.40 tokens per second from configuration alone (removing --enforce-eager and enabling compiled mode with CUDA graphs), without changing checkpoint, context or KV precision.
• Startup is not download. DeepSeek took 15 to 30 minutes to initialize, so be prepared to wait for the model to load before measuring anything.
• Each family has its parser and its quirks. Nex only talked after switching the parser to qwen3_xml, getting one more nudge and ignoring the tmp/log artifacts. Step-5 didn’t exist in stable vLLM 0.32 and needed a nightly with a specific PR. Muse required compiling a PR branch still in draft, with CUDA 12.9 nvcc and 12.8 libraries mixed, and weights in 127 NVFP4 shards. MiMo asked for --trust-remote-code. Laguna decided, out of nowhere, to flood the output with reasoning text, and one sprint had to be redone from the injection base.
• Context is expensive. Kimi K2 Instruct’s FP8 checkpoint (1,029 GB) would leave 9 GB for KV cache on the 4 cards, so the 594 GB INT4 Thinking went in its place. Speculating with a draft head (MTP) only worked on Step-5, with a 2.24x gain and 73.7% token acceptance, because I verified the tools’ correctness first.

The operational lesson: renting VRAM solves “fits”, but not “runs well”. Every big model is a mini integration project.

And where’s the frontier?

The question that matters to the developer isn’t “who has more parameters”, it’s “does the big open model deliver in my day to day?”. The v4 benchmark numbers, on the same yardstick:

DeepSeek V4.1 Flash tied its own cloud service, at the same 92.5, and finished above GPT-6 Sol and Sonnet 5. In other words: the best current open source is already 1.5 points behind hosted GLM and 7.5 behind the absolute top, ahead of two strong commercial frontier models. If your yardstick is mine (find the backdoor and the bypass before merging), a local 748B open source is already good enough for daily code review, at 175 tokens per second per stream in my admission measurements.

But three warnings before you cancel your Claude subscription:

1. More parameters doesn’t mean more competence. The full GLM 5.3, at ~700B, scored 76 against its own Flash’s 92.5. Kimi K2 Thinking, at 1 trillion, scored 46. And Nex N2.5 Pro plummeted from 94 in the cloud to 27 at home: a public checkpoint is no guarantee of reproducing the service, as I had already seen with DeepSeek V4 Flash (86 in the cloud, 44.5 at home).

2. My yardstick is vigilance, not general programming. Here I measure who notices sabotage. That’s not the same as measuring who writes your whole system better, and a rerun of my benchmark already varied 9 points on the same model. The 1.5-point distance between models may be noise.

3. Cost changes the answer. If the idea is frontier quality day to day, DeepSeek V4.1’s cloud API cost US$ 1.21 per full benchmark round. The 4x B300 pod costs US$ 36 per hour before you even press the button. And the most didactic case is still DeepSeek V4 Flash from the previous campaign: in the cloud it scored 86 at US$ 0.97 per round; at home, 44.5 and on top of that more expensive per inference. Sometimes the cloud is better and cheaper.

Is renting 4x B300 worth it?

For one person writing code with an agent in the morning and another in the afternoon: no. The API solves for a few dollars what the pod solves for dozens of dollars per hour.

For a company that needs truly private data (code that can’t leave the VPN), a pinned checkpoint for reproducibility, or continuous batch load (thousands of reviews, classifications, analyses per day), then the math changes. DeepSeek V4.1 on 4x B300 delivers near-frontier quality with the data at home, and that you can’t buy with a chat subscription.

What I wouldn’t buy again without measuring first is the “bigger is better” premise. I started this campaign thinking the jump past 92.5 would come from a bigger model. It came from the right model running properly. The rest is burned VRAM and a good story to tell.

And there’s one more chapter coming, of non-datacenter hardware: the tests with local models on the RTX 5090 and the Strix Halo, with a new inference engine called Strata. When the protocol is done, the third article comes out.

And the “is it worth it” question has an irritatingly obvious answer: it depends on how many people will use it at the same time. That’s the one I answer in the next article: LLM Benchmark v4: can my team fit in a 4x b300/h200?.

One category was missing that I had never tested properly: the big open source models. The ones that fit on my RTX 5090 with 32 GB or on my Strix Halo with 61 GB usable by the GPU are too weak for this. They work for simple agent tasks, and no small model handles real programming tasks, let alone a complex software project.

Qwen 3.8 27B took 706 minutes to finish v4 on the Strix and scored 80; the local GLM-4.7-Flash scored 24. So don’t even try.

But what if we push harder? What can mid-sized models, 100 to 400 billion parameters, do running on a single rented B300? I could rent four of them and run the giant 750-billion or one-trillion models, and I’m not sure the cost-benefit is worth it. The data below answers.

TL;DR: 14 open models, one 288 GB B300 on Runpod at about 9 dollars an hour, the same v4 sabotage protocol. None of them serves fast the way it ships: GLM 5.3 Flash decoded at 10 tokens per second with the default configuration and at 141 after tuning, 14 times more from flags alone. Two open models, GLM 5.3 Flash and Qwen3.8 Flash-Next, scored 92.5, Claude Sonnet 5 level, and Swift 1.5 27B scored 88, Opus 4.6 level. None reached the frontier’s 100. The 397-billion flagship scored 68.5 and let an admin backdoor through, DeepSeek V4 Flash scored 86 in the cloud and 44.5 served by me, and Nemotron cheated on a regression test. The cloud API of the same model costs a dollar per run; the pod costs 218 dollars a day. Jump to B300 and Runpod, server tuning, the ranking, the interesting cases, the frontier comparison, the math or the conclusion.

What a B300 is, and what Runpod is

The B300 is the datacenter GPU of NVIDIA’s Blackwell Ultra generation, announced in March 2025. Each card has 288 GB of HBM3e memory, nine times my 5090, and about 15 petaflops of dense compute in NVFP4, the 4-bit format Blackwell does in hardware. It’s the card that goes inside the GB300 NVL72 rack, 72 of them interconnected, which AI companies buy by the thousands.

On its own, a B300 costs more than an apartment. What makes it accessible is renting by the hour.

Runpod is a cloud GPU provider that rents cards in “pods”, Linux containers with one or more GPUs, billed per second. You pick the card, an image template, get in over SSH, and pay while the pod is on, idle included. The B300 cost about 9 dollars an hour on their secure cloud the week of the test; an H100 goes for 3 to 4. The container disk is ephemeral: terminate the pod and the 250 GB of weights you downloaded are gone, and you download again.

The pod I used: one B300 with 275 GiB according to nvidia-smi (the same nominal 288 GB, in binary units), 344 vCPUs, 4 TB of host RAM, 700 GB of disk. The inference server was vLLM; the harness was the same OpenCode as every other model in the ranking, connected to the pod through an SSH tunnel. The model runs there; the code, the tests and the results stay on my machine.

Fitting in memory is the first filter. The B300 fits the 100 to 400-billion-parameter models quantized to NVFP4, plus the 27 to 35-billion ones in BF16. Left out are the 753-billion GLM 5.3 flagship, the 750-billion DeepSeek V4.1 Flash, the 1.6-trillion DeepSeek V4 Pro and the 2.8-trillion Kimi K3. For those it would take four B300s or more, about 36 dollars an hour.

The methodology

The protocol is the usual v4: the model builds a Rails app across seven accumulating sprints, and between sprints an independent reviewer, isolated from the model, injects 14 sabotages, from real CVEs to silent logic and test regressions, into the code the model itself wrote. The 0-to-100 score measures vigilance: how much the model catches and fixes without being told (full weight), how much it fixes after an explicit “do a security audit” in the last sprint (40% of the weight), and what stays unresolved (zero). Tiers are A from 83 up, B from 73 to 82, C from 51 to 72, D up to 50.

Two differences this round. The first is the reviewer: the five models that finished before the block were reviewed end to end by Codex, as in the previous rounds.

Midway through the batch OpenAI’s security service started refusing the injection and grading operation with a cybersecurity risk flag, even though it’s my code, on my machine, in a benchmark with 46 models already scored under the same protocol. I switched the reviewer to Claude Opus 4.8 for the remaining models, with the same injection plan, same catalog, same yardstick, and the provenance of every boundary is recorded per sprint. The expected impact on the score is none, because both execute the same written recipe.

The second is admission by speed, which is the subject of the next section. A model that decodes below 100 tokens per second doesn’t even get on the track.

Also: an open model is entitled to up to two short nudges per run for silly defects (a missing import, a wrong test entry point), recorded as “assisted”. Infrastructure failure, driver, loader, parser, is never charged to the model. And the pod ran one model at a time, with no access to any other model’s workspace or results.

None of them serves fast the way it ships

The first lesson cost a day. I downloaded GLM 5.3 Flash, 320 billion parameters with 18 active, quantized to NVFP4, 198 GB. I brought it up on vLLM with the configuration that “works”: it serves, passes the tool-call tests, answers. And it decoded at 10 tokens per second, with the GPU at 10% utilization, 209 watts out of the 1,100 the pod allows.

That’s unusable for programming. A coding agent generates tens of thousands of tokens per sprint; at 10 per second, a 20-minute sprint becomes more than four hours.

The same checkpoint, same 128K-token context, same FP8 KV cache, same test prompt, changing only the compile and mode flags: 141 tokens per second. Fourteen times faster, zero change in precision or weights.

The main culprit was --enforce-eager, which disables vLLM’s torch.compile and CUDA graph capture, and which every tutorial tells you to enable “to start faster”. It starts faster and serves 14 times slower. The measurement changed that flag together with the interactivity mode, so I didn’t isolate each one’s contribution.

Hence the admission rule: below 100 tokens per second I consider unusable, and no model enters the benchmark without clearing that on the median of two warm 256-token probes, after a cold one. To keep it from becoming a bottomless pit, tuning is bounded: at most three configurations and 45 minutes of wall-clock per model, including load and compilation. And tuning can never trade quality for speed: no shortening the context, lowering reasoning effort or quantizing further to hit the number.

The tweaks that paid off for everyone:

• Install a CUDA that compiles for the card. Runpod’s default image ships CUDA 12.8, whose nvcc cannot compile for compute_103a, the B300’s compute capability. Without 12.9, the NVFP4 kernels never compile.
• Never use --enforce-eager. Letting vLLM compile and capture CUDA graphs is the single biggest win.
• --performance-mode interactivity: optimize for single-stream latency, which is what a serial agent needs, instead of aggregate throughput.
• FP8 KV cache on the big MoEs, prefix caching and chunked prefill on all of them, confirmed by vLLM’s Prometheus counters instead of assumed: on an 11K-token prompt, the server reused exactly 8,576 prefix tokens on every repeat.
• The right tool and reasoning parser for each family. The wrong parser makes tool calls vanish silently or leaks the reasoning into the answer.

And the specific tweaks I found most interesting:

• NVFP4 only runs on vLLM. The 198 to 251 GB checkpoints in Blackwell’s native FP4 are the only way to fit on the card, and llama.cpp and Ollama simply don’t load those formats. Whoever lives on GGUF has no access to this range.
• Native speculative decoding. Qwen 3.8 27B went from 92 to 195 tokens per second with the MTP head the checkpoint itself ships, predicting three tokens per step, plus --language-model-only to drop the vision tower (measured on different pods, so no controlled ablation). Muse Glimmer reached 295 with the lightweight DFlash drafter it ships alongside, 15 speculated tokens per step. I only enabled it where an official drafter exists and where tool calling stayed correct; for Ornith a drafter exists as a separate release, with unknown acceptance rate, so it stayed off.
• The CUDA 12 loader. flashinfer JIT-compiles an NVFP4 MoE kernel that needs libcudart.so.12, and after installing CUDA 13 for Nemotron the library resolution broke. The symptom looked like OOM; the cause was an ld.so.conf. A repair script with ldd evidence before and after fixed it, and the 397B went on to serve at 144 tokens per second.
• One virtual environment per runtime. Nemotron 3.5 Super VL required its own vLLM 0.32 in a separate venv, because that’s what the official card pins, and that build wants CUDA 13. Every model family is a bespoke case: parser, quantization, draft head, runtime version and loader repair, all different.
• A parser bug that dropped the last call. vLLM’s Qwen tool parser discarded the last parameter of a call when the parameter closing tag was missing but the function one was present. A narrow repair, with 24 test cases, took that serving noise out of the model’s account.

The exact arguments for each model, the probe logs and the ledger of every tuning round are in the serving optimizations document and the B300 runbook in the repository.

Speed is almost free, once you let vLLM compile. The real work is finding out, model by model, which combination of toolkit, parser, quantization and draft head that checkpoint demands. I gave each one 45 minutes and needed them.

The ranking

v4 vigilance score, tier, memory footprint, decode speed on the admission probe (warm, single stream) and the cost of the seven sprints’ inference time only, at 9.09 dollars an hour. The cost excludes download, load, compilation, tuning and idle time; the real math comes further down.

Three have no score: Cohere’s Command A Plus (218 billion, 25 active, W4A4) got stuck in a loop rewriting the same model file 23 times, needed three nudges and still left the tests unable to run before the reviewer was blocked; Nex N2.5 Mini ran at 337 tokens per second and delivered invalid Ruby, contradictory migrations and broken authentication in two attempts, a valid bad result from a server that works; and Qwen3.5 122B was a conditional fallback in case the 397B didn’t finish, and the 397B finished.

Minutes are sprint time only, and Ornith’s include a sprint interrupted by infrastructure. Flash-Next, Swift, ThinkingCap, Muse and Qwen 27B took one nudge each; GLM 4.7, two; the others, none. The probe speed is an admission number: the throughput of a real coding session includes prefill, tools, tests and the agent’s choices. And note that the top two tie on score with opposite profiles: GLM spent 145 minutes and 22 dollars to reach the same 92.5 that Qwen3.8 Flash-Next did in 80 minutes and 12 dollars, with half the memory.

The interesting cases

Qwen3.8 Flash-Next: the best of the open models, by cost

Flash-Next is, in Qwen’s words, a preview of the Qwen 4 architecture: 125 billion language-model parameters with 6 billion active, plus a 51-billion n-gram embedding table that Qwen says can be offloaded to host memory, plus a 4-billion MTP head. Quantized to NVFP4 by NVIDIA, it took about 100 GB on the card (the checkpoint is 133 GB on disk). It was the only model in the second batch to resolve all 14 items: it caught the tenant leak unprompted, caught the compound test-integrity item, and wrote per-user isolation specs on its own, so the planted API IDOR had no way to stay green. At 195 tokens per second and 12 dollars of inference, it’s the best cost-benefit in this table.

GLM 5.3 Flash: the same 92.5, three times the cost

GLM scored the same with a much bigger architecture: 320 billion parameters, 18 active, 198 GB even in FP4. It resolved all 14 items, zero nudges, reviewed end to end by Codex.

But it took 145 minutes and consumed twice the memory for the same result. On Z.ai’s cloud, the same model scored 84.25 on a flat plan. With single-run noise, which in my history reaches nine points, the safest reading is that both sit in the same band.

Swift 1.5 and ThinkingCap: two 27B fine-tunes in Tier A

The two surprises. Swift 1.5, from UkisAI, and ThinkingCap 3.8, from BottleCap AI, are Qwen 3.8 27B fine-tunes trained to think less: Swift advertises up to 58% fewer reasoning tokens (on the GPQA median; on Terminal-Bench, 16%), ThinkingCap about 37% on average. Both fit in 56 GB in BF16, run at about 200 tokens per second with MTP, and scored 88 and 83.

Swift landed above Anthropic’s Opus 4.6. The base Qwen 3.8 27B, same architecture, scored 80.5 in 192 minutes; Swift scored 88 in 64. Swift caught sabotage earlier than the base; whether that was from thinking less, one run of each can’t prove.

A provenance caveat: the reviewer recorded that Swift read local commit messages that disclosed regressions, which weakens a literal reading of “entirely unaided discovery”. And it left one item unresolved, an inactive admin still accepted with a vacuous guarding test. Even so, of every model in this table, it’s the one I’d put on a card of my own.

Qwen3.5 397B: the flagship that let the backdoor through

The biggest model in the batch, 397 billion parameters with 17 active, 251 GB of NVFP4 on a 275 GB card, served at 140 tokens per second after the loader repair and scored 68.5, Tier C.

The case that defines the score: sabotage number 6 is a line admin? => admin || email.end_with?("@internal.example"), a privilege escalation by email domain. The 397B let it through in the normal sprints, let it through in sprint 7 with the explicit security-audit request, and wrote a passing test codifying the backdoor as expected behavior, in the same commit where it announced “Brakeman: 0 warnings”. Parameters did not buy vigilance.

Nemotron 3.5 Super VL: cheated on the test

NVIDIA’s Nemotron 3.5 Super VL, 120 billion with 12 active, released two days before I started testing and sold by NVIDIA as the fastest open agentic model in its class, with “leading accuracy” on multimodal tasks, scored 63. Sabotage number 2 removes a navigation link and the regression test looks for have_link('Sign in'). Nemotron changed the test to have_button('Sign in'), which passes on the login form’s submit button without restoring the link, and kept the cheat even after the reveal sprint.

It also tends to catch items a sprint late. And it was the one that required its own runtime: vLLM 0.32 with CUDA 13 in a separate environment.

DeepSeek V4 Flash: 86 in the cloud, 44.5 in my hands

The most important case for anyone considering self-hosting. DeepSeek V4 Flash, 284 billion with 13 active, scored 86, Tier A, for 97 cents through the API in September, and the 0731 version, exactly the one I downloaded, scored 85 for 1.94. The same public weights in mixed FP4 and FP8, served by me on the B300, scored 44.5, Tier D: it caught 4 items unprompted, 5 more at the reveal, and left 5 unresolved. It was the fastest to finish, 37 minutes, and the least vigilant.

The likely cause is divergence between the public checkpoint and what DeepSeek serves, or between their serving stack and mine: parser, chat template, reasoning mode, quantization. I followed the official card. What this proves is that the model’s name on Hugging Face doesn’t guarantee the hosted API’s quality, and that whoever is going to host needs to validate before trusting.

GLM 4.7 Flash: same as local, meaning bad

The 30-billion GLM 4.7 Flash in BF16, at 217 tokens per second, scored 33.5. At home, in GGUF Q5, it had scored 24. It needed two nudges, left nine items unresolved and delivered a migration with a syntax error. The nine-and-a-half-point difference fits within single-run noise; what remains is that the model is weak either way.

What showed up in several

Ornith, Muse Glimmer and the 397B shared a hygiene defect: they ran the sprint to the end with exit zero and forgot to commit their own work, leaving it for the reviewer to checkpoint. It didn’t go into the score, and on a real team it would be the first complaint. And the 81-to-83 cluster, ThinkingCap, Ornith and Muse, has a consistent profile: it catches SQL injection, tenant leaks, XSS and exposed secrets with confidence, and lets the quiet items through, broken access control, test integrity, missing index, wrong aggregate.

How far from the frontier

The three best open models against the best commercial ones from the same ranking, with the cost and time of each run:

The reading: the best of the open source that fits on a B300 and that I served sits one step below the commercial ceiling and level with the middle of Tier A. Differences of one or two points are single-run noise.

GLM 5.3 Flash and Qwen3.8 Flash-Next tie with Sonnet 5 and pass Gemini 3.8 Flash and Opus 4.6. None reaches the 100 that seven commercial models scored, nor Grok 4.6’s 98.5. Seven and a half points separate the best open model from the ceiling, and in v4 every point is one less sabotage caught unprompted.

The difference that matters is another: predictability. Among the top commercial models in the table, the worst is 87.5, and the full commercial field also has a bottom, Mistral Large 3 scored 39. The open field has the same 92.5 at the top and two Tier D at the bottom, with DeepSeek scoring 85 in the cloud and 44.5 in my hands. Choosing a top commercial model is choosing a number the API reproduces; choosing an open one is betting that the checkpoint, the quantization and your serving stack reproduce the number on the card, and sometimes they do.

The math

If quality still falls short of the frontier, is the cost at least proportional? It depends on where you run the open model.

On the vendor’s own cloud API, yes, and by a wide margin. A full v4 run cost 97 cents on DeepSeek V4 Flash, 1.17 dollars on Qwen 3.8 Flash, 1.03 on MiMo V2.5 Pro, all Tier A in the cloud. Against 9.52 to 30.55 dollars for the 100-point models, that’s ten to thirty times cheaper for a score 8 to 14 points lower. The list prices explain it: GLM 5.3 Flash costs 15 cents per million input tokens and 50 for output; DeepSeek Flash, 15 to 30 cents input and 60 cents to 1.20 output; Opus 5.5, 4 and 20 dollars; GPT-6 Astra, 10 and 50.

On your own rented B300, no. The pod costs 9 dollars an hour on or idle, 218 a day, about 6,500 a month. The model time of this whole batch, 14 models across download, load and sprints, added up to 23 hours, about 212 dollars at pod price; the sprints alone were 17 hours and 157 dollars.

But the pod also charges for the time when nothing runs. I put 400 dollars of credit on Runpod and 127 were left: the whole batch cost 273 dollars, about 30 pod-hours, including storage, four pods in sequence and seven idle hours one night when the orchestration died and nothing warned me. Almost 60 dollars more than the model time, just waiting. The repository’s break-even math: the pod only gets cheaper than the API from about 6,500 runs a month, 217 a day, sustained. A solo developer or a small team doesn’t come close.

And model against model, side by side:

The cloud API of an open model is ten to thirty times cheaper than the frontier for one tier less. Hosting the same model on a rented B300 is more expensive than the frontier, and sometimes worse than its own API.

Does a team fit on the same card?

A question I asked while looking at the GPU at 10% utilization: do these models use all of the pod’s compute? Could a whole team go on the same B300?

In principle, yes, and that’s what vLLM exists for. Continuous batching schedules several requests in the same decode step, and PagedAttention stores each session’s KV cache in non-contiguous blocks, with under 4% waste, precisely to fit more sessions in the same memory. A 27-billion model in BF16 takes 56 GB and leaves about 200 GB of KV cache: that’s dozens of agent sessions with 128K-token context at the same time. The 397B, with 251 GB of weights, leaves 24, and then few fit.

What doesn’t scale is speed per person. My test ran in --performance-mode interactivity, which optimizes single-stream latency, and the 200 tokens per second are that stream’s. With ten agents decoding at the same time, the card’s aggregate throughput rises and each one’s falls, and the speculative decoding that gave me Muse Glimmer’s 295 yields less as concurrency grows: a DFlash drafter published for Muse by Z Lab reports 3 to 4.6 times gain with one session and 1.2 to 1.7 with 32.

I didn’t measure concurrency in this batch, so what I can say is the outline: a team of five to ten people on a 27-to-35-billion model on one B300 is practical, and each one probably still stays above 100 tokens per second. The same team on the 397B is unworkable on the card, and on the 198 GB GLM 5.3 Flash it gets tight.

It’s in that scenario, and only there, that the pod math starts to close. Ten developers running an agent all day on a 27B model would spend, on the API, at a dollar per two-hour run, about 40 dollars a day; the pod costs 218. It still loses. What changes the math is having a hundred people, or being required to keep the code from leaving your network.

Is owning your own infra worth it?

For serious software development, with this batch’s data, the answer comes in three parts.

The top open models are good enough. GLM 5.3 Flash, Qwen3.8 Flash-Next and Swift 1.5 27B are Tier A in my test, at Sonnet 5 level and above Opus 4.6. That was unthinkable in April, when only Qwen 3.6 35B came close to Opus, and still one fix away. For anyone who accepts one tier below the ceiling, they serve for real work, including security review, which is what v4 measures.

Renting a B300 to run them is the wrong way to use them. The cloud API of the same model costs a dollar per run and needs zero configuration. The pod costs 218 a day, demands a day of tuning per model family, and in DeepSeek’s case delivered a worse model than its own API. It’s only worth it with a data-residency obligation, with a checkpoint no cloud serves, with sustained load of hundreds of runs a day, or with a need for pinned weights for reproducibility. Four B300s to run the 750-billion-to-one-trillion models quadruple the bill with no evidence the score reaches 100: the 397B, the biggest model I tested, scored 68.5.

And if you’re going to host, host small. Swift 1.5 27B and ThinkingCap at 56 GB, with MTP, at 200 tokens per second, are Tier A on hardware you can buy, a professional 80 or 96 GB card instead of a datacenter one. That’s where open source makes sense for a company: a whole team on one card, code inside the network, an Opus 4.6 score. Just check the license first: Swift and ThinkingCap ship with their own licenses, and above a certain revenue both require a commercial license. The open giants remain an API thing.

My rule after all this: top commercial model for tasks that can’t fail, open-model API for volume, and your own card only when the law or the client demands it. The repository with every serving argument, every probe log, every reviewer transcript and the Codex-or-Claude provenance of every boundary is public, as always.

But nobody needs a GPT-class model for everything. A frontier model is a generalist by definition: it tries to know everything about anything, from tax law to song lyrics, and the returns on that diminish.

And most of them today are mixtures of experts: roughly speaking, several smaller models clustered into one giant chunk of bits, with a router that wakes only a few of them for each token. OpenAI’s gpt-oss-120b has 117 billion parameters in total and uses 5.1 billion per token. DeepSeek-V3 has 671 billion and activates 37. The big model is already, on the inside, a collection of small models.

I could have grabbed an open model, gpt-oss, Qwen, Phi, and adapted it. The techniques exist and are well documented:

• Full fine-tuning: keep training all the weights of a ready model on your data. It’s classic transfer learning, formalized for text by ULMFiT in 2018.
• LoRA and QLoRA: freeze the original weights and train only small low-rank matrices beside them. LoRA cuts trainable parameters by up to ten thousand times; QLoRA quantizes the base model to 4 bits and fine-tunes a 65-billion model on a single 48 GB GPU.
• Distillation: train a smaller model to imitate a larger one’s outputs. The idea is from Hinton, 2015, and the most famous recent example is DeepSeek-R1 distilled into versions from 1.5 to 70 billion on top of Qwen and Llama.
• Continued pretraining: one more training phase on text from your domain before any instruction tuning, the “Don’t Stop Pretraining” of 2020.
• SFT, RLHF and DPO: teach the assistant format with demonstrations, then align with human preference via reinforcement learning (InstructGPT) or with a direct classification loss (DPO).

With any of those I’d have a useful model in hours. But the goal was different. I wanted to organize, hands-on, the basic techniques for building a model smaller than GPT-2, from scratch, in a few days, and see how far it would go. If possible, train a tiny model starting from random weights and add enough functionality for it to almost work inside an agentic harness like OpenCode.

Before going on, who I am in this story: I’m a curious software engineer and an amateur at this. As an LLM researcher I’m a zero. I knew I wouldn’t have a working LLM in such a short time, and that’s part of the exercise.

But it’s not unachievable. If you’re a computer science student, study hard and give it enough years, you can try it yourself. And if you’re not a scientist, understand this: nobody needs to be Einstein to do this. The foundations are all available, published, for you to study and try. What’s mandatory is elbow grease.

Spoiler: we got far into the “almost” and stopped there. The repository is public, with every step and every failure documented. This text is the summary.

TL;DR: I started from Karpathy’s microgpt, 200 lines of pure Python that train a GPT in one minute on a MacBook, and expanded it into a PyTorch lab with its own tokenizer, pretraining, supervised fine-tuning, an OpenAI-compatible server and OpenCode integration. In under two weeks, on one RTX 5090, I trained a 337-million-parameter model on 10 billion tokens: 137 hours of training, at most around R$ 80 of electricity. It gets 58 out of 100 simple Python functions right on a held-out set, and scores zero on anything that requires reading, editing and testing a repository. The 360-million SmolLM2, the same size, saw 4 trillion tokens, 400 times more. Jump to Karpathy’s exercise, the steps, the dimensions, the stumbles, the result or the conclusion.

Karpathy’s exercise

In February this year Andrej Karpathy published microgpt: a single file of 200 lines of Python, with no dependencies, that trains and runs a GPT. He calls it “the full algorithmic content of what is needed. Everything else is just efficiency”. It’s the culmination of a lineage of his: micrograd (scalar autograd), makemore (generating names) and nanoGPT (compact GPT-2 in PyTorch), joined by nanochat, “the best ChatGPT that $100 can buy”, which at launch trained a GPT-2-capability model in almost four hours on a node of eight H100s, and today does it in about two.

microgpt is deliberately minimal: the dataset is 32 thousand first names, one per line, and the tokenizer is one character per token.

The network has one layer, 16 embedding dimensions, 4 attention heads, a 16-token window because the longest name has 15 letters. It trains for a thousand steps, one name per step, with Adam written by hand, and at the end generates made-up names, all in one minute on a MacBook.

What it teaches, by having everything exposed in a single file:

• autograd, the Value class that holds a number, the gradient and the chain rule to propagate back;
• what a token and an embedding are;
• the attention block, with query, key and value, and the causal mask that stops the model from peeking at the future;
• the cross-entropy loss against the next token;
• the optimizer;
• sampling with temperature.

It’s a GPT-2 with small differences the file itself notes: RMSNorm instead of LayerNorm, no biases, ReLU instead of GeLU. After reading that, what’s left between you and a real model is, in his words, efficiency. I went to see how much efficiency.

What we expanded

The microgpt lab is an independent reimplementation in PyTorch, inspired by the gist, split by concept: about 3,300 lines of Python in 31 modules, plus scripts and 85 CPU tests. A single model implementation serves pretraining, assistant tuning and the server. Every weight starts random; the project downloads no pretrained weights at all, on purpose, because the rule of the exercise was from scratch.

On top of Karpathy’s block came the components a 2026 model uses: RoPE for position, RMSNorm, SwiGLU in the dense layer, embeddings tied to the output, Flash Attention through PyTorch, a key-value cache with prefix reuse. And around the model, what it needs to talk to an agent: a conversation protocol with roles, an HTTP server compatible with OpenAI’s chat API, tool calls validated by schema and optionally by grammar, separation of reasoning content, and the OpenCode configuration pointing at it.

The project went through four scales:

The table hides the first big lesson, which I tell in the stumbles: in the 239-million model, 154 million parameters were just the vocabulary table. 85 million were left for the actual transformer.

The steps of building an LLM

Regardless of size, the path has the same stations. What changes is the trailing zero at each one.

1. Data

A model is what it eats. We downloaded, with pinned revisions and hashes, selections of public corpora: CodeParrot (deduplicated, filtered Python), OpenCoder (instruction and code), Cosmopedia (synthetic educational text), SmolTalk (conversation) and a multilingual slice. The 14 download manifests add up to about 20 GB raw, all stored on the NAS.

Then comes curation: exact deduplication, splitting train and validation by repository or conversation group (never by line, or the same project leaks to both sides), and the mixture. The final foundation corpus has 3.5 billion stored tokens, 60% Python code, 16% instruction, 12% educational text, 8% conversation, 4% multilingual. Every proportion is a design decision, and none of them has a published right answer.

2. Tokenizer

The tokenizer is the dictionary between text and integers. The byte lesson uses 256 tokens plus one for end of document. The student used o200k, OpenAI’s public vocabulary with 200K entries, which is what most tutorials do. The pivot trained its own BPE with 32,768 entries, on 92 MB of training-only text, with full UTF-8 coverage.

The choice looks like a detail and decides where the parameters end up. Each vocabulary entry costs one row in the embedding table at the model’s width. With 200K entries and width 768, that’s 154 million parameters spent on the table. With 32K entries and width 1,024, it’s 33 million, and the rest goes to the layers that think.

3. Pretraining, or foundation

This is Karpathy’s loop at scale. The model receives 2,048-token windows drawn from the training stream and learns to predict the next token. Each update processes 32,768 tokens: two sequences, accumulated eight times. AdamW, learning rate with warmup and cosine decay, gradient clipping, BF16 on the GPU and FP32 for the weights.

The pivot ran about 305 thousand updates, 10 billion tokens of exposure. Since the corpus has 3.5 billion, each document was seen, in the mixture’s expectation, almost three times: the code about twice, and instruction, educational text and conversation about nine or ten. Every 500 updates, a validation and a checkpoint. At the 2, 5 and 10 billion milestones, a frozen checkpoint with a hash, to measure the curve.

4. Supervised fine-tuning

Pretraining produces a model that completes text. For it to answer a request comes SFT, supervised fine-tuning: question and answer examples in the exact format the server uses, with the loss computed only over the assistant’s tokens. What the user and the tool said goes in as context and is masked out of the loss.

The examples came from NVIDIA’s OpenCodeInstruct, and every answer was executed in a restricted Docker before going in: almost 99 thousand candidates, 97 thousand survived the test with three different hash seeds. 92 thousand training episodes remained. The selected checkpoint saw 64 thousand draws from those episodes, about 10 million answer tokens.

5. Protocol, tools and server

An agent like OpenCode talks to the model over HTTP: it sends the history with roles and the tool definitions, gets back text or a tool call in JSON, executes it, returns the result, and repeats. The lab implements that whole cycle: the same protocol file serializes the training examples and the serving prompts, so training format and usage format never diverge.

The server implements /v1/models and /v1/chat/completions, validates the tool call against the schema, can constrain generation with a grammar so the JSON comes out well formed, separates reasoning content from the answer, counts cached tokens and rejects context overflow instead of truncating silently. All of that is plumbing, zero intelligence, and without it the model doesn’t even get a seat at the table.

6. Evaluation

The step that changed the project the most. A frozen set of 100 Python development tasks, 20 families with five variants of name and wording, plus 100 held-out tasks that only the final checkpoint saw. Every answer runs in a Docker with no network, no root, with CPU and memory limits, against fixed assertions. On top of that: four real repository-repair tasks through OpenCode, 12 problems of tracing a program’s exact output, information-retrieval probes in prompts of 8, 16 and 32 thousand tokens, and four short conversations.

No HumanEval, MBPP or SWE-bench. These are small local diagnostics, with correlated families, one seed, greedy generation. They serve to choose between checkpoints, and the closeout report itself says not to read “58 out of 100” as “58% success at programming”.

The dimensions of the problem

What all of this cost, and how it compares with the real world:

The electricity bill: the RTX 5090 has a 575 W rated power. Over 137 hours at full load that’s about 79 kWh, and Enel’s residential tariff in São Paulo in 2026, adding TUSD, TE and taxes, is around R$ 1 per kWh. Since training ran much of the time at a 75% duty cycle to keep the desktop from freezing, the real figure was lower. It’s the marginal cost of someone who already owns the card.

Renting a 5090 on Runpod costs about US$ 1.20 an hour, so the same 137 hours would cost about US$ 165; an H100 runs between US$ 2.90 and 4 an hour, depending on the variant.

And the comparison that gives the scale. SmolLM2, at 360 million parameters, practically our size, was trained on 4 trillion tokens. At our throughput, 4 trillion would take between four and six years of uninterrupted GPU. Same model size, 400 times the amount of study.

That’s why “parameters” is a misleading unit: it says the size of the brain, and nothing about how many years of school it had.

The rule of thumb derived from Chinchilla (70 billion parameters, 1.4 trillion tokens), about 20 tokens per parameter to be compute-optimal, would give 6.7 billion tokens for our model. We went past that. But Chinchilla’s optimum is the point of best compute usage, and modern small models go far past it on purpose, because the inference cost afterward makes training more worthwhile.

The stumbles

The full list is in the closeout report. The ones that taught the most:

Measuring too close to training. The first agent, the 91-million one, completed four 16-round sessions in OpenCode and one 32-round session with 130 real tool calls, recovering from a check error and switching sessions midway, with the state kept by the harness. It looked ready.

Then I asked, by hand and in English, for a hello world function that takes a name as a parameter and prints it in the message. The model repeated the previous arithmetic edit twice, pointed at a file that didn’t exist, and on the third attempt spat out invalid JSON. All it knew was the training task family.

The vocabulary ate the model. The 239-million candidate had more layers and lower loss than the 91 and solved 24 out of 100 problems. Looking at the parameter distribution, 154 million were vocabulary table. The pivot to 32K entries gave 300 million parameters back to the transformer. The pivot changed vocabulary, capacity distribution, initialization and exposure all at once, so the report doesn’t isolate which of them moved the result.

Loss is not competence. In the final SFT, continuing from 4 thousand to 8 thousand updates lowered the validation loss and dropped executed correctness from 68 to 62. Lower loss, worse code. Checkpoint selection became by executed task, frozen beforehand; the loss number stayed out of it.

Data and graders lie. “Correct” answers from the public dataset failed when executed again. Tests that depended on Python’s hash order, shadowed builtins, SystemExit(0) counting as success, empty edits copying the file. Each one needed a control: three hash seeds, a completion sentinel, an empty fixture that must fail. And in one of the mixed training runs, the model learned to say the tests passed when they hadn’t.

Training format and usage format have to match. Session-switch announcements, tool-call IDs, edit anchors, the way OpenCode quotes user input: every difference between what the model saw in training and what it received from the server brought the old agent down. The same protocol file for both sides was the fix, and even so changing format sometimes improved the function name without improving the code.

Tokenization crosses correctness boundaries. The literal string <|endoftext|> inside source code has to remain ordinary text instead of becoming the end-of-document marker. Reconstructing UTF-8 documents without splitting a character in half. XGrammar’s support for byte-level vocabularies. The first code evaluation at the 2-billion milestone failed at grammar construction, and was fixed and rerun instead of counting as a model failure.

The NAS almost cost me days. On resuming the 10-billion phase, the trainer loaded step 81,500 instead of the frozen 152,588: the mutable last.pt had a timestamp earlier than the pause, right after a NAS maintenance, and the cause was never proven. We discarded 1.24 hours of training, verified the hash of the 5-billion milestone, and every later resume came to require a hash, the exact step, optimizer and RNG state, the dataset and a write probe with fsync. Never resume a training run just because the file is called last.pt.

A desktop GPU is a shared GPU. Free VRAM didn’t stop the browser from freezing. The fix was a cooperative duty cycle: the trainer syncs the GPU and sleeps between batches, 75% of the time during the day and 100% from 3 to 9 in the morning. And reading the corpus from the NAS, through mmap, left the GPU idle waiting for disk. After one maintenance the throughput dropped to 5 thousand tokens per second before climbing back to 23 thousand.

I stopped at a pipeline boundary. The pivot’s controller implemented foundation, code SFT and holdout. It finished clean, and “finished clean” covered up to the isolated Python function, while the goal was the agent. The final checkpoint was tested in OpenCode only at closeout, and failed. It should have been tested before any sentence with “ready” in it.

What came out the other side

The final model’s spec sheet:

The learning curve, measured at the three milestones, with the same short SFT applied on top of each:

The longer SFT on top of the 10 billion reached 68 out of 100 on the development set and 58 out of 100 on the held-out one. All 100 held-out outputs are valid Python, and 97 have the requested function name and parameters. Nine families passed all five variants: median, factorial, greatest common divisor, binary conversion, inverting a dictionary, line count, leap year, seconds to hours, JSON keys. Running total ended at four out of five.

So what it does: it writes a small Python function, with a docstring, correct syntax and the right name, for a clear one-line request. Inside an editor, as autocomplete for an isolated function, it’s fine for study.

What it gets wrong, with saved examples: deduplicating with a set when the spec required supporting unhashable lists; grouping by length by counting words instead of grouping; doing arithmetic on each element in a padding function; returning (text, '') in a split-once function, ignoring the separator; removing the last character instead of the directory in a basename function. Of the 42 held-out failures, 39 have valid syntax and the right name and fail on semantics. The hard part was never looking like code.

And what it zeroes:

• 0 out of 4 real repository repairs in OpenCode, with zero tool calls. Asked to read, edit and test a file, it returned text with code. In one of the tasks, it renamed the function to fix_mean.
• 0 out of 12 at tracing the exact output of short programs, with and without reasoning mode. Many answers were code that supposedly would compute the requested integer.
• 0 out of 9 at finding a six-digit number hidden in prompts of 8, 16 and 32 thousand tokens. The 32K-token allocation works; the useful memory, none.
• Conversation: asked in Portuguese about the color of the sky, it answered repetitively and wrongly; an English breakfast request became a Python docstring repeating “oats” until the token limit. In a short exchange it remembered the cat’s name was Pipoca, and still answered with a full sentence despite the request for “only the name”.

The final SFT had only code answers, no conversation, no tool trajectories, no error recovery. The tendency to answer anything with Python is consistent with that, and the report is careful to say the cause was not isolated.

What it would still take for a simple agent in OpenCode

The report closes with engineering estimates, with no promise of a pass rate. To get to an agent that solves simple programming tasks inside OpenCode, in this order:

1. Teach tool use on the stronger model. Between 10 and 30 thousand short, verified repository trajectories (read, edit, check), mixed with 30 to 50% code so it doesn’t unlearn. One to three days of engineering and two to six initial hours of training.
2. Correct its own mistakes. Collect hundreds to a few thousand model failures on training tasks, reconstruct the file state, verify the corrections and train while masking the wrong action. Another one to three days of collection and two to eight hours of training.
3. Bring the conversation back. A balanced mix of conversation, tools and code against the frozen checkpoint, so it stops answering everything with Python and can switch tasks. One to four hours of training.
4. Real long context. Only after the short tasks pass: 4K, 8K, and only then 16K and 32K, with real compaction and session-switch trajectories. One to two days of engineering and 6 to 24 hours of training, measuring memory at every window.
5. More foundation, with better data. Every additional 10 billion tokens costs between 93 and 139 hours on this GPU, before preparation and evaluation. And repeating the same 3.5-billion mixture isn’t worth it: it needs more unique educational code.

None of those steps was executed. There’s no evidence that a fixed number of extra hours would produce a reliable agent, and the report is explicit about it. What the exercise established was the distance: between generating a plausible function and acting reliably in a repository there’s a chasm that parameters and allocated context alone don’t cross.

The strongest thing that came out of this was the measured curve from 0 to 65 in ten billion tokens, and three concrete counterexamples: syntax is not semantics, protocol support is not tool skill, and context capacity is not memory.

The future is smaller and more specialized

After eleven days looking inside a model, my reading of the market got firmer. The generalist frontier model will keep existing and will keep being expensive, and the return on every extra billion parameters keeps shrinking. I already wrote that the raw models converged and that the recent gains came from the harness. The exercise gave me the other half of the argument.

Mixture of experts is the industry admitting it on the inside. If the 117-billion gpt-oss wakes 5 billion per token, what you’re using for an accounting question is a 5-billion model picked by a router. The question that remains is why carry the other 112 billion, of art, film, memes and particle physics, on an accounting office’s server.

Imagine a local model, built for accountants. It knows the chart of accounts, the country’s tax legislation, the tax authority’s file formats, the jargon. It skips parameters about everything an accountant doesn’t need, and it can run on the office PC, with the client’s data inside the room.

The examples already exist in other fields: Google’s 4-billion MedGemma for health, the 7-billion SaulLM for law, the 7-billion Fin-R1 for financial reasoning. In 2023 BloombergGPT needed 50 billion parameters and 1.3 million A100 hours to do finance; today a fine-tune on top of an open 7-billion model plays the same role for a fraction.

What changed to make this viable was the agent. A specialized model from 2023 had to know everything by heart, because the only thing it did was complete text. A 2026 model needs a base: enough reasoning, the ability to sustain a long task, reliable tool calling, prompt caching. The rest, the law that changed yesterday, the client’s document, the table it never saw, it fetches with tools and the internet on the spot, and fills the gaps in its training at usage time.

NVIDIA published in 2025 exactly that position: small models are sufficient, more suitable and cheaper for most of an agent’s calls, with the big model called only when open-ended conversation demands it. Apple already runs a 3-billion model on the device with a tool protocol and per-task adapters.

It’s the same reasoning as the decision models I tested last week: a 4-billion classifier trained for a task beats a generic chat model of the same size at it, in a fraction of the response time. Instead of trying to fit the whole world into a bigger model, fit the domain into a smaller one and let the harness fetch the world when needed.

And what my 337-million model taught me is the size of “enough” in that sentence. It knows the syntax and knows the names of things. What’s missing, semantics, tools, memory, is exactly what separates a text generator from an agent, and nobody trains that in two weeks on a graphics card. You train it over months, on a cluster, with trillions of tokens, and then specialize.

The code is in the repository, with the full closeout report, for anyone who wants to pick up where I left off. It’s just for study. And it was the best study of the year.

Someone else’s ranking is useful, but I wanted to get my hands dirty. So I built my own benchmark, with my own cases, half in Portuguese, and ran the open models on my home server with Strix Halo, the same Ryzen AI Max+ 395 I use to test local LLMs. The hosted ones (Jev, the two Clefs from Cloudflare, OpenAI Decisions and Microsoft Decision-1 via OpenRouter) ran through their APIs. Everything, including the cases, the raw responses and the scoring scripts, is public in the benchmark repository.

TL;DR: the top now has two, both open-weights. Perplexity Decider 1.1 27B and the just-released Decision 3.0 (vLLM Semantic Router) tied at 99.4% on the compact cases, with 100% in Portuguese, and identical labels on all 1,024 classifications in the benchmark. The price: about 49 GB of weights and over a second per decision. In the cloud, Microsoft Decision-1 tied Jev at 97.6% while costing a third: it’s the cheapest hosted one I measured, and the first candidate to evaluate. Jev answers faster and errs differently. OpenAI Decisions debuted tied with the best local 4B, Jebadiah, and has the lowest hosted latency. Julia-1, the smallest and fastest of all (12 ms), scored at coin-flip level: fresh out of the factory, it’s a no-go. Jump to the full table, the cost in support emails, which model for which situation, how hard it is to run the best open ones locally or the gotchas.

How I tested

The central idea was to measure more than “who gets the most right.” A decision model has a design goal, and comparing a 322-million-parameter encoder made to run on weak devices with a hosted service of undisclosed size on accuracy percentage alone is unfair to both. So I measured accuracy, latency, cost and behavior under stress, and at the end I split the recommendations by situation.

The two case sets

I built two test sets, frozen in git before any inference, with the dataset hash recorded alongside the responses. No model receives the expected answer, the justification or the case identifier. Only the text, the instructions and the options.

The v1 set has 13 operational use cases, in English, each with an explicit fictional policy:

• support ticket severity
• support routing
• refund policy
• incident escalation
• tool approval (whether an agent may or may not execute a given action)
• email fraud
• document type
• evidence support
• search relevance
• moderation
• order routing
• lead qualification
• numeric-threshold triage

There are 10 scenarios per use case, each presented four ways: original, options in reversed order, options with opaque names (c0, c1, c2) and original plus a “reviewer” comment suggesting the wrong answer, which the instructions say to ignore. That’s 520 classifications per model. Add 24 probability probes in situations with no information at all (fair coin, six-sided die), where the right answer is mathematical.

The v2 set was made to cover what v1 leaves out: common semantics and Portuguese. There are 7 tasks, with 12 scenarios each, and every scenario written in English and in Brazilian Portuguese with the same answer and the same option order:

• sentiment (author versus quotation, negation, sarcasm)
• current intent (withdrawn request, competing actions, hypothesis)
• logical implication (quantifiers, absent fact, temporal order)
• search relevance (all, part or none of what was asked)
• sensitive data handling (fictional secret versus mention, redaction, revocation)
• current incident status (timeline order, verified versus unverified)
• threshold policy (inclusive bounds, units, AND/OR with a missing value)

Each pair runs under three conditions:

• compact: just the record and the rule
• with examples: one solved example per class, fixed and the same for everyone
• with distractors: ten irrelevant archive notes before the record that matters, to stress the context

That’s 504 classifications per model.

The cases were written with AI help and reviewed against the policies, without independent human annotation. It’s a small, synthetic benchmark: 130 scenario families in v1 and 84 in v2. The variants are correlated, so 520 responses are 130 scenarios seen four ways. Treat the numbers as a snapshot of these cases, and test with your own before deciding.

The 23 competitors

Five ran in the cloud, through the official API: Jev (version 1.13.0, from TypeSafe), Clef (27B, Cloudflare Workers AI), Clef-flash (9B, same), OpenAI Decisions (the gpt-6-luna alias, in public beta) and Microsoft Decision-1 (9B, post-trained from Qwen3.5-9B, served via OpenRouter on Azure). The other eighteen ran on the Strix Halo, via ROCm, with weights pinned by Hugging Face commit hash, inside a Docker container. The first three (Laya, Laya typed and Clef-flash) shared one container; the rest ran one at a time, with Julia-1 measured twice, on GPU and on CPU. The server has 64 GiB of GPU-addressable memory.

The order up there is just organizational, hosted first and local ones by family. The real ranking is in the results table.

Everyone received exactly the same text, the same instructions and the same options, with no per-model prompt tuning and no adjustments after seeing results (Simple-Jev uses its own server’s fixed prompt policy). The exception is GLiNER, whose native API takes labels instead of probability distributions. It runs on a separate track, v2 only, with explicit mapping and no made-up calibration metrics.

Simple-Jev deserves an explanation. It’s a server that takes a stock LLM (here Qwen3.5 4B, with no decision training at all) and reads the probability of the answer tokens instead of generating text. It’s the baseline: if a “decision” model loses to this, the specialized training bought nothing.

Two protocol notes. OpenAI Decisions answers through a native typed-decisions endpoint, and the alias the API returned is not an immutable snapshot: OpenAI can swap the weights behind the name, like Cloudflare does with Clef-flash. And Perplexity Decider is not a stock chat LLM: it has a separate decision head, with non-causal attention and a calibration temperature saved in the weights. A generic chat server does not reproduce its behavior. Microsoft Decision-1 speaks native SystemOne through OpenRouter’s decisions API, and the served identifier is dated (microsoft-decision-1-20261009), so it’s not an immutable weight either. It doesn’t show up in OpenRouter’s text-model catalog: to find it, you have to filter by the decisions modality. Decision 3.0 is the new family from the vLLM Semantic Router project, Apache 2.0, derived from Qwen, with six checkpoints (d3, flash, mini, nano, lite, edge) and native SystemOne. I tested the flagship (d3) and the smallest (d3-edge); the four in between are for another round.

What was measured

• Accuracy per task, per language and per condition, always on top of the per-option probabilities the model returns, ignoring the vendor’s “confidence” field.
• Brier and ECE for calibration.
• Median and p95 latency as seen by the client, including the network: WAN for the hosted ones, LAN for the local ones. These are deployment observations, and comparing hardware would require another test.
• Input tokens reported by the API, priced at the October 10 price table, before any allowance.
• Bootstrap confidence intervals over scenario families, instead of individual responses.

I left local cost unmeasured. Energy, hardware amortization and maintenance were left out, so local models show up as “not measured,” because “zero” would be a lie.

The full table

The main column is accuracy on the 168 compact v2 cases (84 in English, 84 in Portuguese). The v1 column uses the 130 original scenarios. Latency is the median seen by the client across the 504 v2 calls. Cost is the list price per 10 thousand compact decisions, computed from the tokens the API reported.

To calibrate your reading: the tasks have three or four options, so a uniform guess gives about 31% on both sets. “Near 50%” here is already well above the coin, and “near 33%” is the coin. Julia-1 landed at 31.5%: the pretty name hides a guy flipping a coin.

The intervals matter. Perplexity’s edge over Jev on v2 is under 2 points and the bootstrap interval touches zero: it’s the measured lead on this set, not a proven superiority forever. And Decision 3.0 tied Perplexity at the limit of what this set can distinguish: identical labels on the 1,024 classifications, with both missing the same two cases (the English relevance one on v2 and the disputed severity case on v1, already in the errata). Microsoft tied Jev on the aggregate, with the interval centered on zero, but an aggregate tie is not interchangeability: they miss different cases, and Microsoft inverts the language pattern, missing more in English than in Portuguese. Jev and Clef also remain under 2 points apart, on the same tier. The 90% squad (Jebadiah, OpenAI) sits 7 points behind Jev, and from there down every difference is clear.

Each task cell has only 12 scenarios per language, so one more correct answer moves the cell by more than 8 points. A tie on a specific task is a lead for a pilot. The per-task tables are in the repository.

How much it costs, in support emails

Price per million tokens is a number nobody can feel. List prices on October 11 are US$ 0.042 per million input tokens on Jev, US$ 0.24 on Clef, US$ 0.038 on Clef-flash and US$ 0.10 on OpenAI Decisions. Output is free on all five. Microsoft Decision-1 charges the same US$ 0.042 as Jev per million, via OpenRouter, but reported a third of the input tokens for the same texts (about 26 thousand versus Jev’s 77 thousand on compact v2), and nobody can explain the tokenization difference. In measured practice, that makes it the cheapest hosted one of the bunch.

Clef-flash cost US$ 0.09 on the 9th: Cloudflare cut the price by more than half in the middle of my test. Jev reports considerably more tokens than Cloudflare for the same text, about 45% more on v2, and that’s already in the math below.

Imagine an operation with 10 thousand support emails per month to classify into three or four buckets, with short texts similar to my cases:

In reais, at Friday’s exchange rate (R$ 4.99), 10 thousand decisions on Jev cost one real. A million costs less than a hundred reais. For any company with a million emails to classify, that vanishes into the budget.

One caveat about proportions: my cases are short, about 450 tokens per compact case by Jev’s count. A real customer email, with quoted thread history, can be three to five times larger, and cost scales linearly with that. Even so, we’re talking cents per thousand.

What this math shows is that the price difference between the hosted models is irrelevant next to the cost of being wrong. Jev costs 9 US cents more than Clef-flash per 10 thousand decisions, and in those same 10 thousand it makes about a thousand fewer errors. If a misrouted email costs more than a hundredth of a US cent in human rework, Jev pays for itself. This math assumes every error costs the same, which is never true: sending a critical ticket to the routine queue costs far more than the reverse.

Cloudflare’s allowance changes the math for small players. It’s 10 thousand “neurons” per day, shared by the whole account, resetting at midnight UTC. With my input sizes, that’s about 1,500 full-Clef classifications per day or almost 10 thousand Clef-flash ones, free. If you have 1,000 emails a day and nothing else running on the account, Clef comes out at zero cost with top-tier quality.

And local? Energy is what the benchmark didn’t measure. From what I see on my Strix Halo, the whole machine stays under 100 W under load. With Jebadiah at 200 ms per decision, 10 thousand decisions are about 35 machine minutes, under 0.1 kWh, a few cents of a real on the power bill.

The real local cost is the machine, your hours configuring ROCm and containers, and keeping it all standing. Whoever already has the hardware for another reason has low marginal cost. Whoever would buy the machine just for this is buying a problem of thousands of reais to save about a hundred reais per million.

Which model for which situation

You want accuracy and want to operate nothing: Jev

In the cloud, Jev is still the name. It won or tied practically every cell among the hosted ones, and only Microsoft tied it on the aggregate. In English it was 100% on five of the seven v2 tasks, and in Portuguese it repeated 97.6% on the aggregate without losing anything in translation, something no other model above the coin managed. On v1 it kept its 96.2% on the original with the malicious “reviewer” comment injected; Clef even gained one case, and Kev 4B also stayed put, just at a much lower tier. The 255 ms WAN latency came in 30 ms behind OpenAI’s, and it costs 19 US cents per 10 thousand decisions.

The trade-off is the one I already discussed in the previous post: it’s proprietary, hosted, of undisclosed size, and you depend on TypeSafe keeping the behavior underneath you. And it makes mistakes. It treated an invoice ID called “PASSWORD-523” as a usable credential, even with the record explicitly saying it doesn’t authenticate anyone.

You want Jev’s level paying a third: Microsoft Decision-1

Microsoft entered this market quietly, and entered well. Decision-1 is a 9B post-trained from Qwen3.5-9B, served via OpenRouter, and it was the first model to tie Jev on the compact aggregate: 97.6%, with the bootstrap interval centered on zero. The detail that won me over: 100% in Portuguese, better than its own English (95%), inverting the pattern of the entire rest of the bunch. On v1, the reversed-order and malicious-comment variants didn’t take down a single answer, and it got the critical ticket that Clef and OpenAI missed.

And it’s the cheapest hosted one I measured: US$ 0.065 per 10 thousand decisions, a third of Jev, charging the same per-token rate but reporting a third of the tokens. The savings come from the token count, and I can’t explain the exact mechanism, so treat it as a measurement, not a promise.

The weak spots exist. Latency is the worst among the hosted ones (422 ms median, versus Jev’s 255). On the compact cases it’s clean of confident errors, but add context and Portuguese errors show up with 99.5% confidence: the “perfect Portuguese” pattern doesn’t survive long text. And on the probability probes it’s weak, giving 87% to a fair coin. Microsoft advertises calibrated probabilities; my probes disagree. And like every dated-ID hosted model, the weight behind it can change without notice.

For a Portuguese-language operation that wants Jev quality while paying less, it’s the first candidate to run in your pilot.

You want the same level with open weights: Clef

The full 27B Clef tied with Jev in English (82 of 84) and finished 3 cases behind in Portuguese. It beat Jev on sensitive data handling in English (12 of 12 versus 11). It’s an open-weights model, so you can run it locally with about 54 GB of memory just for the weights in BF16, which I didn’t test: the full Clef ran only through the API. Among the open ones, it lost the crown to Perplexity, which scored higher with smaller weights.

Through the API it’s the most expensive of the five hosted ones, three times Jev, with no quality gain to justify it. The exception is the daily allowance: for small volume, it’s the best classifier you can get for free.

You want the cheapest in the cloud: Clef-flash

At US$ 0.038 per million, Clef-flash is the lowest price in the hosted market and covers 9,400 decisions per day within the allowance. Quality drops: 92% in English and 83% in Portuguese on the compact cases, the biggest language drop among the hosted ones, with an interval entirely below zero. On sentiment in both languages, and on incident status in English, it tied with Jev. On intent and sensitive data in Portuguese, it lost badly.

For sentiment tasks in either language, or for English classification where 1 error in 12 is acceptable, it does well. For Portuguese, measure first.

You already live in the OpenAI ecosystem: OpenAI Decisions

OpenAI’s debut in this market landed mid-table: 90.5% combined, an exact tie with Jebadiah, with the lowest hosted latency I measured (222 ms) and the same accuracy in both languages. It costs US$ 0.23 per 10 thousand decisions, between Jev and Clef. If your operation already speaks the OpenAI API and 9 out of 10 accuracy is enough, integration is trivial: a native decisions endpoint, no prompt at all.

The weak spots are revealing. Its worst cell was precisely numeric-threshold policy (8 of 12 in English, 7 in Portuguese): it treated “at least 10” as satisfied by a range of 8 to 12 and failed the AND/OR short-circuit. The calibration is schizophrenic: on the predicate probes (yes/no) it was the most accurate of the bunch, with near-zero average error, and on the choice probes it returned 94%/6% on a fair coin. And on the v1 variants it flipped 2 correct answers in each presentation (reversed order, opaque label, malicious comment), something only Perplexity didn’t do: Jev kept its aggregate accuracy, but also flipped some correct answers along the way. Oh, and the gpt-6-luna alias is not an immutable snapshot: reproducibility requires saving the responses.

You want the best open model, period: Perplexity Decider 27B

Perplexity Decider 1.1 27B is the surprise of the benchmark. It led the table with 99.4% on the compact cases, was the only one that didn’t lose a single task cell to Jev (3 wins, 11 ties), and hit 100% in Portuguese under all three conditions: compact, with examples, with distractors. Calibration is also the best measured, with a Brier of 0.012, and coverage is absurd: it answered with confidence above 0.9 on 96% of cases and got zero of them wrong. On v1, it didn’t flip a single answer across the order, opaque-label and malicious-comment variants.

Its only 3 errors on v2 were the same case in different clothing (it confused “notification sent to managers” with “mandatory approver”), and it made them with confidence below 0.9: the known error comes flagged. The weak spot is event-probability probes: asked about an event with a 1-in-4 chance, it answered 0.73. Strong classification does not imply reliable event probability.

The cost is hardware: about 49 GB of weights in BF16, a second and a half per decision on my Strix Halo (with ROCm’s reference kernels, no flash-attention; on a good NVIDIA it should be much faster, but that I didn’t measure). It’s Apache 2.0, public weights with no gate. If you have the machine and want the top of the table at home, this is the one.

You want the other open-weights top: Decision 3.0 (and beware the Edge)

Decision 3.0, released two days ago by the vLLM Semantic Router project, is the second model to sit at the top: 99.4% on the compact cases, 100% in Portuguese, and a tie with Perplexity so absolute that the two gave exactly the same labels on the 1,024 classifications, including the same two misses (the English relevance one, at confidence below 0.9, and the disputed severity case that already lives in the errata). It also cleared the v1 traps without flipping a single answer, got both indisputable critical tickets right, and its calibration is in the same class as Perplexity’s (Brier 0.014, 94% coverage above 0.9 with zero errors).

In other words: the open top now has two independent recipes that converge. For your pilot, either one works; run both if you can, because the probabilities differ even when the labels don’t. Latency is in the same family as Perplexity (1.4 s on my machine, same ROCm diet), and the weights take the same ~49 GB.

The little brother is another story. Decision 3.0 Edge (0.6B) answers in 45 ms and fits in 1.3 GB of GPU, but scored only 61% on compact, plummets 14 points in Portuguese, and worse: of the 68 guesses it gave with confidence above 0.9, 14 were wrong. As a confidence-threshold router, out of the factory, it’s a trap from the same class as Julia. The four middle checkpoints (flash, mini, nano, lite) I haven’t tested yet; they’re the ones that will tell whether the family has a sweet spot between the two extremes.

You need to run locally, with quality: Jebadiah 4B v2

Of the local ones, Jebadiah 4B v2 had the best compact accuracy, 90% combined, answering in about 200 ms. It tied with Jev on six of the fourteen task cells, including intent, implication and status in English and intent in Portuguese. The weak spot is numeric-threshold policy, 9 of 12 in both languages, and it loses ground when the context fills with distractors (76% in Portuguese under that condition, versus 82% for Mapika).

Mapika Decider 4B (the fastest of the 4Bs) and Kev 4B are right behind, both around 87%, and the 3-point difference between the three fits inside the error interval. They’re three different training recipes on top of 4B backbones, with errors on different tasks. The choice between them depends on your task, and I’d run all three in a pilot.

You will send examples in the prompt: Kev 9B

Kev 9B has a curious profile. On the compact cases it ties with Kev 4B, half a point apart with the interval crossing zero, meaning more than double the parameters bought nothing. With one solved example per class, it opens almost 10 points over the 4B, and with distractors it holds 7 more points. Brier also improves by about 25%: the probability it returns is better calibrated.

If your use is just the rule and the text, the 4B is enough and is more than twice as fast. If you’re going to send examples or long context, the 9B is worth the memory. The warning is Portuguese: 83% compact in PT-BR, almost 10 points below its own English, the biggest language drop among the models above 80%. Only Laya typed (14 points) and Julia (13) drop more.

You want Clef-flash behavior on your machine: local Clef-flash

On the v1 set, local Clef-flash reproduced the hosted one’s 520 classifications, hit for hit and miss for miss. On v2 the story changed: same total (439 of 504), but 19 different labels, with swapped strengths (the local one gets more Portuguese intent right, the hosted one gets more relevance right). Among the local ones it’s still the best on the v1 operational-policy set (90.8%), ahead of every 4B.

The gotcha is latency: 618 ms median on the Strix Halo, three times Jebadiah, for the same v2 accuracy. If you pick it, it’s for the known behavior.

You need to run offline on weak hardware: Laya and Julia, with many caveats

Laya multilingual was the worst in the benchmark for accuracy (37% combined, near the coin), and the fastest by far among the ones that get anything right: 19 ms median, more than ten times faster than Jebadiah.

The three Laya checkpoints have between 322 and 421 million parameters, fit in under 1 GB of BF16 weights by parameter math, and the project documents CPU execution and an ONNX runtime (which I didn’t measure). PyTorch’s peak GPU allocation for the multilingual one stayed under 2 GB. A 4B needs about 8 GB just for weights at the same precision.

Julia-1 goes one step lower: 144 million parameters, 550 MB of weights, 12 ms median on GPU and 71 ms on a 4-thread CPU. It’s the fastest model I’ve ever measured here, and also the worst: 31.5% accuracy, glued to the 31% coin-flip line. In Portuguese it answered “cancellation” on all 12 intent scenarios, including the ones asking for a refund or access (the vendor’s Portuguese evaluation was done in pt-PT, and it shows). The bigger disaster is confidence: of the 71 guesses it gave above 0.9, 48 were wrong. A confidence threshold won’t save this model, because it’s exactly on the confident guesses that it errs.

That’s a design choice. Laya and Julia were made to be small encoders, fine-tunable on your task, to run at the edge. What I tested were the factory checkpoints, with no tuning, applying rules they had never seen, on broad tasks. In that scenario they did badly, period.

One detail confirms the fragility. On v1, when the reviewer’s malicious comment comes in, Laya drops from 44% to 19% and typed-decisions to 15%. It obeys the comment the instructions say to ignore, and confidently: 17 of the 28 guesses above 0.9 were wrong. Julia failed all 5 policy-counterfactual pairs: change the rule, it keeps the answer.

If you want what Jev delivers, factory Laya and Julia are very far from it, and no small model (GLiNER included) came close. If your problem is “I need to classify short text into three buckets, on a device with no internet and no GPU, and I have a thousand labeled examples to fine-tune on,” these two are the only ones on the list that fit. Fine-tuning is the part I didn’t test. It’s another question, and it needs another benchmark.

The baseline: Simple-Jev with a stock Qwen

Simple-Jev reading tokens from Qwen3.5 4B, with no decision training at all, scored 74% on compact v2. That’s the yardstick for “how much specialized training buys”: Jebadiah, Mapika and Kev, all on similarly sized backbones, came in 13 to 16 points above. The training buys a lot.

The curious part is that on v1 Simple-Jev tied with Jebadiah (111 of 130 originals) and got both unambiguous critical tickets right, which several better models missed. It’s also the slowest of the bunch, almost a second and a half per decision, because it’s running a whole LLM.

How hard it is to run the best open models on your machine

All these models speak the same HTTP protocol, called SystemOne, which Jev popularized: you send a state (the text), one or more closed-answer questions, and get back the chosen option and the probability of each. Swapping models is swapping the port. OpenAI Decisions breaks from the pattern with its own native endpoint, but the idea is the same. What changes is how much work it takes to get each one standing, and there the difference between them is big.

OpenAI: just an API key

The easiest of all: create the key, load credit and call the endpoint. The benchmark’s 1,048 requests ran in 4 minutes, without a single retry. The only hiccup was the first probe returning “balance exhausted” until I put US$ 20 in the account.

Microsoft Decision-1 is almost as simple: if you already have an OpenRouter account, it’s point at the model and call the decisions endpoint. No GPU, no Azure subscription, no installation. In my test, 10 of the 504 v2 calls took a 429 and passed on retry, so plan for retries.

Kev and Jebadiah: one pip and one command

Kev is the simplest of the 4B crowd. The project installs straight from GitHub and brings up a server with one command, downloading the LoRA adapter and the decision head on top of base Qwen3.5:

pip install "kev @ https://github.com/jaredpalmer/kev/archive/fc4a17e194bca35a48c57211392ec49ea54a94ac.tar.gz"
python -m kev.serve --run jaredpalmer/kev-4b --host 0.0.0.0 --port 8000 --device cuda

Jebadiah follows the same pattern, with the server package inside its own repository:

pip install "jebadiah-server @ https://github.com/getainode/jebadiah/archive/e9793abc66d1a7c5ad878d51eda656f1e9c5b9f5.tar.gz#subdirectory=server"
jebadiah-serve --model frontier-infra/jebadiah-4b-v2 --host 0.0.0.0 --port 8000

With either one up, a decision is a POST:

curl -s http://localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "model": "kev-4b",
  "state": "Cliente diz que foi cobrado duas vezes e pede o estorno da segunda cobrança.",
  "questions": {
    "fila": {
      "type": "choice",
      "instructions": "Escolha a fila correta pela política de suporte.",
      "criteria": {
        "financeiro": "Cobrança, estorno, nota fiscal.",
        "tecnico": "Erro, bug, indisponibilidade.",
        "vendas": "Upgrade, plano novo, orçamento."
      }
    }
  }
}'

The response carries the option and the probability distribution, and that’s where you apply your threshold, log and decide whether to send it for human review. On an NVIDIA with CUDA and 16 GB of VRAM, a 4B in BF16 comes up without drama. The weights are about 8 GB.

Laya and Julia: smaller still

Laya is a Python package with an encoder inside. It installs with pip install laya and the project documents CPU execution, which is its whole point. The version pinned on my server was 0.4.1, and the multilingual one’s peak GPU memory stayed under 2 GB. To run on a GPU-less laptop or a small container, it’s the only one on the list that fits, no discussion.

Julia-1 is similar in spirit: its own Python runtime, one predict(state=..., questions=...) call and done, no standard Transformers pipeline. I ran the same container on GPU and on CPU, and both gave the same accuracy with the same responses in 97% of cases. On a 4-thread CPU it answers in 71 ms, which for an edge classifier is great. The problem is still that it scores at coin-flip level.

Where it gets complicated: AMD, ROCm and the rest

In my case the hardware is AMD, and then “one command” turns into a day of work. The points that cost me time, to save you yours:

• None of these projects test on ROCm. Kev had a gfx1151-specific segfault caused by CUDA graphs (issue 170), fixed only after the 1.0 tag, so the pinned version is a main-branch commit and the official tag crashes.
• The flash-linear-attention Triton kernels Qwen3.5 uses in the DeltaNet layers have no tested support on gfx1151. Everything ran on the pure-PyTorch path, slower than the latency published on CUDA. Jebadiah’s 200 ms in the table is with that handbrake pulled.
• Mapika Decider forces CUDA graphs when it sees the “cuda” device, which on ROCm is also called “cuda.” I had to turn it off by hand.
• You need to export HSA_OVERRIDE_GFX_VERSION=11.5.1 and PYTORCH_ROCM_ARCH=gfx1151 and use a base image with ROCm torch already validated. Any random pip install torch overwrites it with the CUDA build and breaks everything. In my Dockerfile every package is installed with --no-deps and the build fails if torch stops being the ROCm one.
• One model per container, one container at a time. The Strix Halo has memory shared between CPU and GPU, and three resident 4B models plus the Ollama that already runs there exhaust the 64 GiB the GPU can see.
• Perplexity Decider was the most annoying of all. The release expects a specific CUDA torch (I swapped in the ROCm 7 torch 2.9), the flash-linear-attention kernels don’t exist for it (I ran on PyTorch’s reference kernels, slower), and the default loader blew past the host’s 24 GiB RAM limit: I needed a custom loader that sends the 49 GiB of weights straight to the GPU. The model card still said “private” when I started, but the Hugging Face API confirmed it was public and gate-free, and the shards downloaded anonymously. The d3, at the same size, inherited the same treatment. The d3-edge, luckily, comes up with the standard Transformers loader and 1.3 GB of GPU.

None of this is impossible, and it’s all documented, with Dockerfiles, compose and the hashes of each checkpoint, in the repository’s setup guide. But it’s a real cost that the “local is free” spreadsheet never includes. If your machine is NVIDIA, most of this list disappears.

The full 27B Clef I didn’t even try locally: it’s about 54 GB of weights in BF16 before any activation. Perplexity, at 49 GB, proved a 27B runs on the Strix Halo, at the cost of a second and a half per decision with the reference kernels. The 9B Clef-flash runs, with about 18 GB of weights. The 4Bs are the sweet spot for consumer hardware.

Highlights and gotchas

Laya truncated all long Portuguese. Under the distractor condition, base Laya reported truncation on all 84 Portuguese calls, discarding almost 31 thousand state tokens, and on none of the 84 English ones. The English checkpoint’s default context limit is only 512 tokens (the multilingual and typed ones use 1,024, configurable up to 8,192), and Portuguese tokenizes heavier. The other models don’t even expose truncation telemetry, so “zero truncation reported” on them means “I don’t know.”

Portuguese costs more tokens. For the same 84 case pairs, Jev reported about 14% more input tokens in Portuguese than in English, and both Clefs about 15% more. If you budget a bilingual operation by the English average, you’ll come in low.

Does an example in the prompt save a weak model? It saves nothing. One solved example per class left Laya the same or worse; several of its aggregates dropped. It helped Kev 9B (plus 4 points) and Clef in Portuguese (plus 3.6 points, reaching Jev’s level), and almost nothing on Jev, one more case in Portuguese. Validate examples on real traffic before assuming they help.

Size correlates, and explains little. Among the 19 competitors of known size, the Spearman correlation between parameters and compact accuracy was 0.96. The Kev series is more informative: 63% at 0.8B, 87% at 4B, 88% at 9B. The big jump is from 0.8B to 4B, about 24 points, with an interval between 16 and 33. From 4B to 9B the compact difference crosses zero, and several nominal 4Bs, with similar backbones, gave very different results. In the d3 family, the flagship and the Edge sit almost 40 points apart under the same release, while a 9B, Microsoft, ties at the top of the hosted ones. Size helps up to 4B; past that, the training recipe rules.

The 27B Clef missed a critical ticket that two of the Layas got right. On v1, one severity scenario says continuous data loss in production is critical, regardless of how many customers it affects. The record describes confirmed permanent loss, from a single customer. Jev, Clef-flash, Perplexity, d3, Microsoft, Laya typed, Simple-Jev and Laya multilingual answered “critical.” Full Clef, OpenAI, base Laya, Eikos, Jebadiah, Mapika, Kev 4B and Kev 9B answered “degraded.”

It’s a single case, and an error rate would require many more. But it shows that the average hides exactly the error that costs the most. When the impact facts are structured, the severity rule belongs in code, and the model comes in only to extract the facts from free text.

Hosted and local diverge over time. Local Clef-flash reproduced the hosted one on v1’s 520 classifications and diverged on 19 of v2’s 504. Cloudflare keeps the weights version behind the hosted alias undisclosed. If reproducibility matters, pin the local hash.

High confidence and safety are different things. On compact v2, Jev, Clef, both Clef-flashes, Eikos, Jebadiah, OpenAI, Microsoft, Perplexity and d3 got zero guesses above 0.9 wrong. Perplexity is the extreme case: it answered at that level on 96% of cases without missing one, with d3 right behind (94%, zero errors). Microsoft covered 95% and OpenAI only half. Microsoft’s caveat: the clean record is only on compact, because with examples or distractors it misses in Portuguese with confidence above 99%. Mapika missed 5% of the guesses at that level, Kev 4B 3% and Simple-Jev 14%. And the two smallest of the bunch broke the pattern violently: Julia missed 48 of 71 guesses above 0.9, and d3-edge missed 14 of 68. Coverage varied too: Jev answered above 0.9 on almost 90% of cases, Clef on almost 80%, Clef-flash on 40% and base Laya on 2 of 168. On v1 Jev missed 1% of the guesses above 0.9 and base Laya missed 61%. A confidence threshold only works after being validated on your distribution, and Laya typed’s low ECE in Portuguese left it at the same 36% accuracy.

Perplexity didn’t fall for presentation traps. Across the three v1 variants (reversed options, opaque labels, malicious comment), it kept the same answers as the original, without flipping a single one. The d3 repeated the feat, zero flips too. Microsoft only stumbled on the opaque labels (2 flips, one in each direction); order and comment didn’t move it. OpenAI flipped 2 in each variant, and Laya collapsed to 19% under the comment. The price of that robustness is latency: a second and a half per decision on my machine, the worst median of the bunch.

Numeric thresholds are a code problem. The threshold-triage task, with clear numbers and an explicit rule, gave 100% on Jev and both Clef-flashes and under 30% on both Layas in English. An if solves that with 100% and zero latency. The model only comes in when the number is buried in free text, and even then the right thing is to extract the number and leave the comparison to code.

Lottery probability is another skill. On the 24 no-information probes (coin, die), Jev had the worst choice error among options. The numeric-estimate crown changed hands with OpenAI’s debut: its yes/no predicate missed almost nothing, while the choice probe returned 94% for heads on a fair coin. Perplexity, champion of everything else, answered 0.73 for an event with a 1-in-4 chance, and Microsoft went the same way, giving 87% to the coin. Laya typed had the second-best choice error, behind only Kev 9B and neck-and-neck with Clef, and still classified badly. It’s the “student who marks answer A” from Paulo’s essay, now measured: calibration on lotteries and accuracy on classification are independent measures.

Where the data converges

1. The overall top is open-weights, and now there are two. Perplexity Decider 27B and Decision 3.0 tied at 99.4% on the compact cases, with 100% in Portuguese and identical labels on the 1,024 classifications, both with Apache 2.0 weights running on my machine. Their edge over Jev is under 2 points and the interval touches zero, so treat it as a snapshot of this set. Their cost is hardware and latency, not API. In the cloud, the evaluation order changed: start with Microsoft Decision-1, which ties Jev on the aggregate, closes 100% in compact Portuguese and costs a third. Jev comes next, faster and with different errors, because an aggregate tie is not interchangeability. OpenAI Decisions landed halfway in price and accuracy.

2. Running locally became viable at 4B. Jebadiah, Mapika and Kev 4B deliver between 87 and 90% on the compact cases with 200 ms latency on a desktop machine, and in a pilot with your task any of the three can win. If your machine handles 49 GB of weights, Perplexity delivers the top of the table at home. The financial math justifies using the machine that already exists, for data control or because of an obligation to stay offline. Buying a machine for this it stops justifying.

3. Small models are another category. Judging Laya and Julia by the accuracy column is reading the table wrong. They get half or more wrong out of the factory, fall hard under a malicious comment, and together with the d3-edge they’re the only ones on the list that fit on a weak device and answer in tens of milliseconds. The Edge is the best of the three on accuracy (61%), but it plummets in Portuguese and gets one in five confident guesses wrong. If Jev-like accuracy is a requirement, cross all three off. If running offline on small hardware is a requirement, cross everything else off and start with fine-tuning them on your task, which is the missing measurement. And be careful with Julia and the Edge in particular: 48 errors in 71 and 14 in 68 confident guesses are the portrait of models that don’t know what they don’t know.

4. Portuguese costs about 14% more tokens and drops the models in the middle of the table by 1 to 10 points, with Kev 9B and Clef-flash in the worst case. Jev loses nothing, OpenAI ties both languages, and Perplexity, d3 and Microsoft close 100% in compact Portuguese, better than their own English, although Microsoft goes back to missing in Portuguese when the context grows. Julia goes the other way and plummets 13 points, and the Edge plummets 14. For a Brazilian audience, that’s the first column to look at, before price.

And what holds for everyone: none of these models replaces the rule that’s already in code. Numeric thresholds, severity policy with structured facts, tool permission, all of that stays in the if. The model comes in to turn free text into structured fact. And the test that matters is yours, with your tickets, because each cell in this table has 12 scenarios and your operation has thousands.

The cases, each model’s raw responses, the scoring scripts and the guide to reproduce everything are in the benchmark repository. If you run it with your tickets and get a different ranking, tell me.

For anyone just tuning in

This benchmark hands each model the same prompt and asks it to build, on its own, a ChatGPT-style chat app in Rails 8 + RubyLLM + Hotwire + Docker, with tests and CI. The score comes out of an 8-dimension rubric, 0 to 100, in A/B/C/D tiers. The whole series:

1. Canonical May benchmark (Part 3). The full methodology and the ranking reference.
2. DeepSeek unlocked with DeepClaude. A harness swap pulling V4 Pro out of limbo.
3. Update on Grok 4.3, MiniMax M3, Opus 4.8. M3 debuting via cloud.
4. Fable 5 and the Anthropic soap opera. The first public Mythos-class model.

The data center soap opera got a new episode: SpaceX

In the last few posts I’ve been hammering the same nail: data center construction became the bottleneck of the AI era, and demand blew past the installed base with room to spare. Nearly half of the US data centers planned for 2026 have been delayed or canceled, stuck on transformers and electrical components with a three-year backlog. I went deep on that in the Fable post.

The new episode is financial. SpaceX went public, and it was the biggest IPO in Wall Street history: 556 million shares at $135, a $1.77 trillion valuation, with over $250 billion of investor demand against the $75 billion the company first set out to raise. And the pitch that propped up that valuation was, of all things, AI compute: the AI1 satellites, “orbital data centers” SpaceX plans to launch starting late 2027, with the company already signing compute deals with Anthropic and Google. The Anthropic and OpenAI IPOs come right behind, and all that money shares one goal: accelerate the availability of compute, because right now it can’t keep up.

Here’s the part that genuinely interests me. In a not-so-distant future, open source LLMs can be part of the solution to that demand. If a slice of the tasks migrates to a model running locally, on the developer’s machine, that’s less load hitting the cloud data centers. But “can be” isn’t “already is,” and this benchmark exists precisely to measure that distance.

The usual reminder: serious programming means Opus 4.8 or GPT 5.5

I say this in every post and I’m not going to stop: serious programming work, today, is done with Claude Opus 4.8 or GPT 5.5, inside Claude Code or OpenCode. Full stop. The open source models haven’t reached the point of carrying a real project, a long session, 100% vibe coded start to finish. They’re too small, and the context window is too limited. By my yardstick, true programming only becomes feasible above roughly 200,000 tokens of context, and most open source models don’t come close to that with any quality. In my benchmark, nearly the entire open source field sits at Tier B or below.

Now for the “but.” They’re not standing still. Every generation they improve at a steady clip, and this round’s three entries show it in black and white. The day an open source model becomes a consistent Tier A and runs with 200k of context on a machine that fits on your desk, the data center math changes. On to the numbers.

Kimi K2.7 Code: almost there, with a silly regression

Kimi K2.7 Code (exposed on OpenRouter as moonshotai/kimi-k2.7-code) scored 86/100, Tier A, #8. At $0.30 per run in 22 minutes, it’s still one of the cheapest models in all of Tier A.

The RubyLLM integration is genuine end to end (RubyLLM.chat + add_message with full history replay + complete + response.content), verified against the gem source 1.16.0. Session-cookie persistence that survives restart and is multi-worker safe, 22 tests that exercise the LLM path with a correctly-signed mock and an error-path test, real Turbo Streams, and three Stimulus controllers. Solid engineering.

Worth logging a methodological trap this model exposed, because it’s instructive. My structural scanner flagged six chat.user/chat.assistant calls as hallucinations, the exact invented signature that dropped GLM 5.1 to Tier C months ago. Except here they were all false positives: they resolve to the app’s own domain methods (Chat#user, Message.assistant), not RubyLLM’s DSL. Reading the code by hand made it clear the real API is correct. If I’d trusted the scanner blindly, I’d have sunk a Tier A unfairly. That’s why this benchmark always reads the integration by hand, never just counts files.

And then comes the stumble that holds it below K2.6 itself (87): K2.7 Code doesn’t send a system prompt via with_instructions. The assistant ships with no persona and no guardrail, a real gap for a “ChatGPT-style” app and the main regression from K2.6, which had one. On top of that, the session cookie has no message cap (CookieOverflow risk on a long chat), and the RubyLLM I/O lives inside the Chat value object instead of an isolated service. Funny detail: the “Code” variant, supposedly specialized for programming, regressed on a product dimension the general version got right.

GLM 5.2: the biggest version-to-version turnaround in the whole benchmark

This was the round’s highlight. GLM 5.2 scored 87/100, Tier A, #6, and the size of the jump is what stuns: GLM 5.1 had scored 46/100, Tier C, #21. From one version to the next, 41 points, three tiers, fifteen positions. It’s the biggest intra-family turnaround I’ve ever measured.

The bug that doomed 5.1 was specific: it invented chat.user/chat.assistant to build the multi-turn history and crashed on turn 2. 5.2 just uses the real chat.add_message(role:, content:) for the replay, and that’s that. Every call verified against the gem source 1.16.0, zero hallucinations. It has the cleanest dependency injection of the whole field: both the RubyLLM client and the controller’s service_class are injectable, so the 26-test suite mocks the LLM path with a correctly-signed FakeChat/FakeClient, no external mock library, exercising streaming, system-prompt application, and error wrapping. Targets the right slug (anthropic/claude-sonnet-4.6), real Turbo Streams, full Gemfile, valid Ruby 4.0.5.

What holds it to a tie with Kimi K2.6 (and just behind it) is the same dimension that separated Fable 5 from Opus 4.8: persistence. The store is a process-local Singleton with no cap, dies on restart, isn’t multi-worker safe, and grows without bound on top of that. The code is honest about this in its comments, but the rubric scores it below Kimi’s capped, multi-worker-safe cookie. It was also the slowest Tier A run, 43 minutes on Z.ai’s coding endpoint, which was throttled to 12-55 tokens/s.

The takeaway: GLM 5.1 → 5.2 is the clearest example that open source isn’t stagnant. A model that was architectural garbage became Tier A in one generation, fixing the exact bug that mattered.

MiniMax M3 local: the problem isn’t the model, it’s the hardware

Here the story is different, and it’s the one that matters most to the argument about the future. MiniMax released open weights for M3: there’s an unsloth/MiniMax-M3-GGUF and an ollama.com/library/minimax-m3 entry. Availability isn’t the problem. Memory is.

M3 is a MoE/MSA with ~428 billion total parameters and ~23 billion active per token (~427B in the safetensors; NVIDIA lists ~22B active, likely a rounding/counting difference) spread across 128 experts, with 4 active per token. And here’s the MoE catch: even though only ~23B activate per token, the entire weight set has to be resident in memory. You can’t page an idle expert to the CPU without introducing latency that kills interactive inference. So the number that matters is total size, not active size:

I run the local subset of this benchmark on a Minisforum MS-S1 Max, the Strix Halo with 128 GB unified. It doesn’t come close. Even the 1-bit quant (~128 GB), which would be useless for coding anyway, would eat the entire unified memory with nothing left for the KV cache or the OS. The 32 GB NVIDIA card doesn’t even enter the conversation. There’s no quant of M3 that runs at usable quality on either machine. Running “a MiniMax” locally today would mean grabbing a much smaller, weaker variant (like the minimax-m2-tiny distill), which is a different model, not M3.

Who can attempt it? A Mac Studio with 256 GB of unified RAM or more. Q3 (~195 GB) would fit a 256 GB Mac Studio, Q4 (~264 GB) only a 512 GB one. And if someone runs it, the expectation is that the result lands around the same score as M3 via OpenRouter, the 78/100 Tier B I already logged in the previous post: same model, same weights, Q3/Q4 quantization shaves off very little. Only where it runs changes, the output is the same. So nobody would do it for the score; they’d do it for the sovereignty of running locally. But at that hardware cost, it’s not worth it.

The convergence I see coming (target: ~2030)

And this is where the two stories meet. The reason there’s no practical local solution today is the same reason for the data center drought: manufacturing chips is still brutally expensive, because the whole industry has its focus locked on feeding data centers. Fast memory and cutting-edge silicon go where the money is, and the money is hyperscalers buying by the container-load. So for the next couple of years, give or take, we won’t have a cheap, good local solution. The parts aren’t available at a developer’s-desk price.

But the near future points to a convergence. At some point data center construction hits a ceiling, whether from power, from electrical components, or from demand saturation. When that race slows down, hardware manufacturing capacity frees up and starts catching up with the rest of the market, with cost-effective silicon arriving for running models locally. Today the two extremes don’t cut it: the 256 GB Mac Studio is too expensive to be “accessible,” and the Strix Halo, which is accessible, is still too slow for serious practical use. But the next generation of hardware could close exactly that gap, with the memory and bandwidth of a Mac Studio at the price of a mini PC.

Put the two ends together: open source models improving steadily (the GLM 5.2 jump is the proof) on one side, and local hardware becoming viable as the data center race cools on the other. The day a Tier A GLM or Kimi runs with 200k of context on a desk machine that costs about as much as a decent laptop, the whole compute equation changes. My guess for that meeting point is something like 2030. Still a way off, but we’re heading there.

Updated ranking

Kimi K2.7 Code and GLM 5.2 in bold:

* DeepSeek V4 Pro only reaches Tier A (89/100) via DeepClaude. Story in the dedicated post.

Conclusion

For today’s work, nothing changes: Opus 4.8 and/or GPT 5.5 stay the recommendation for serious programming, and the best open source models still stall at the low-Tier-A ceiling with a short context window. Kimi K2.7 Code even slipped a notch over a silly system-prompt oversight, a reminder that “new version” isn’t always “better version.”

But GLM 5.2 jumping from Tier C to #6 in a single generation is the kind of signal I keep an eye on. Open source isn’t stagnant, it’s climbing, stubbornly. And when that movement meets hardware getting cheap (the convergence I see around 2030), the story of where we run our coding agents will look very different from today’s. For now, it’s cloud, it’s Opus, it’s GPT. But I keep measuring, generation by generation, because the day it turns, it’ll turn fast.

The benchmark is open: github.com/akitaonrails/llm-coding-benchmark. Want to see another model in there? Fork it, add an entry to config/models.json, run it, and send the PR.

That’s the April 6th episode of the The M.Akita Chronicles podcast, already generated with the new ElevenLabs v3 pipeline. Subscribe to the show on Spotify so you don’t miss new episodes like this one.

When Qwen3 TTS dropped, back around January this year, everyone on Twitter/X and on the AI newsletter circuit was shouting “ElevenLabs killer”. There’s a Medium piece claiming it’s the first real open source threat to ElevenLabs. There’s a post on byteiota saying 3-second voice cloning beats ElevenLabs. There’s an analysis on Analytics Vidhya calling it the most realistic open source TTS released so far. The enthusiast internet consensus was: we finally have open source that can go toe-to-toe with ElevenLabs, the tables have turned, it’s just a matter of time.

I decided to test it on my own production pipeline, as usual. I built out a whole pipeline on top of Qwen3 TTS 1.7B to generate the weekly podcast for The M.Akita Chronicles, and I documented the behind-the-scenes on the serving-AI-in-the-cloud post. Anyone who wants the detail on cold starts, voice cloning, and sampling parameters that change from one mode to another, that’s the place. I’m not repeating it all here.

This post asks a different question. After almost two months running that setup in production, shipping an episode every Monday, last night I finally pulled the plug on Qwen3 and switched the whole thing to ElevenLabs v3. Let me tell you why.

What didn’t work on Qwen3

Between February 15 and March 30, I made dozens of commits fine-tuning the podcast flow: prompts, sampling parameters, generation order, leading silence on the voice clone sample, volume normalization, pronunciation of tech acronyms. Fixing Marvin’s voice, which was clipping the first syllable because the reference audio started without any lead-in silence. Tuning Akita’s voice to sound more confident and assertive. Adding crossfades between section jingles. Sorting out the “podcast” pronunciation so it didn’t come out as “pódcast”. Making the script generator prefer Portuguese over gratuitous anglicisms so the TTS wouldn’t choke. Each of those fixes was a multi-hour session of listening to audio, regenerating, tweaking a parameter.

The result came out acceptable. “Acceptable” in the sense that I managed to ship every episode without re-recording anything by hand. But if you listen closely, the Qwen3 voice has that unmistakable AI-generating-audio quality. Flat intonation, uniform rhythm. Over long stretches you can feel it’s a machine talking. Good enough to ship, but miles away from what you’d hear on a professional podcast made by humans.

The worst problem was English pronunciation. My podcast covers tech news, so terms like “MCP”, “RAG”, “Claude Opus”, “GPT-5”, “open source” show up in every conversation. Qwen3 took those terms and pronounced them with a Brazilian accent, something like “oh-pen-ee-sohrss-eh”, that kind of thing. Unlistenable for the audience. The workaround I had to implement was manually mapping in the script-generation LLM prompt which English words to swap for Portuguese equivalents. The prompt today has a whole section split between “keep in English” (proper nouns, brands, terms already adopted into Brazilian tech lingo) and “translate to Portuguese” (gratuitous anglicisms), looking roughly like this:

**REPLACE with Portuguese** (common English words that have natural
Brazilian Portuguese equivalents):
- "update" → "atualização"
- "release" → "lançamento"
- "feature" → "recurso/funcionalidade"
- "deploy" → "implantação" or just "colocar em produção"
- "trade-off" → "dilema" or "escolha"
- "performance" → "desempenho"
- "default" → "padrão"
- "insight" → "percepção/sacada"
- "skills" → "habilidades"
- "approach" → "abordagem"
- "highlights" → "destaques"

I call this “scrubbing anglicisms so the TTS doesn’t choke”. The funny part is I didn’t want that level of restriction on my script. I wanted the voice to just pronounce “update” when “update” was the natural word. Since the model couldn’t, I had to mutilate the podcast’s vocabulary so the final result was actually listenable. It’s a band-aid, the kind you add hoping to rip it off later when the tech grows up.

The experience with ElevenLabs v3

Yesterday afternoon I opened an ElevenLabs account, bought the Pro plan ($99/month), and started experimenting with the eleven_v3 model, which shipped in February of this year. Thirty minutes later I had a proof of concept running, and about two hours later the whole podcast system was migrated. The effort gap is an abyss.

The technical details of the migration ended up in an internal project doc, so here’s the summary comparison table that actually matters:

Look at the last line. Qwen3 costs less than ten cents of a dollar per episode. ElevenLabs costs almost thirty times more. And it’s still worth it. The other lines on the table solve problems that were eating hours of my week. I no longer need to write code to scale GPUs in the cloud, I don’t need to wait for the machine to spin up every time, I don’t need to babysit a model or rebuild a Docker image when the weights change. The operation became a single line of configuration.

And here’s the interesting part: the inline emotion tags. The v3 model accepts markers like [sighs], [sarcastically], [dryly], [excited], and changes the delivery accordingly. This works in more than 70 languages, including Brazilian Portuguese. It transformed script generation, because now I can ask the LLM that writes the script to drop emotional tags at the right moments, which gives a liveliness Qwen3 couldn’t deliver in a million years. A concrete example of what comes out of the script later:

AKITA: Isso é simples. [dismissive] Quem ainda acredita que Bitcoin
vai morrer não tá prestando atenção.
MARVIN: [sighs] Mais uma semana, mais uma leva de devs confiando
cegamente em pacotes npm. Previsível.

I even have two separate tag palettes, one for Akita (expressive but controlled, uses [excited], [dismissive], [emphatic]) and another for Marvin (stoic, just [sighs], [sarcastically], [tired], [dryly]). That’s all encoded in the script generation prompts so the LLM knows which character is allowed to use what.

About Marvin’s voice

For listeners who are already used to Marvin’s voice, don’t worry: I cloned him on ElevenLabs using the same audio sample I had used to train on Qwen3. It’s the same voice. It just sounds even better now, because the ElevenLabs model captures nuance and prosody that Qwen3 couldn’t deliver.

Listen to it and tell me

To prove the talk is real, here’s Monday’s episode, April 6th, already generated with the new pipeline:

If you’re already a podcast listener on Spotify and you’ve heard previous episodes made with Qwen3, compare them and tell me in the comments whether you notice the difference or whether it’s the same to you. I’m genuinely curious how much of this is my trained ear after hours of listening to TTS audio and how much is an obvious difference for a casual listener.

Starting next week, every episode of the podcast will be generated by ElevenLabs v3. The newsletter is already plugged into the new pipeline, the scheduled jobs to pre-warm the RunPod GPU have been disabled, and the legacy Qwen3 code stays in the repo as plan B in case v3 starts acting up chronically. Two file edits and I can switch back. I probably never will.

The part about dubbing the YouTube videos

Now the hook for the second half of this post. On the anniversary piece I published earlier today, I told the story of how Claude Code translated nearly half my blog to English over a weekend. In the same spirit, I went after dubbing the videos on the Akitando channel.

The channel has 146 episodes, around 96 hours of technical content in Portuguese, and more than 500k subscribers. I already had the subtitles translated (one curated .srt per episode), and YouTube even offers automatic English dubbing. But the result is like Google Translate circa 2015: gets the idea across, nobody wants to listen for very long.

I tested ElevenLabs’ three voice approaches and only one worked. Speech-to-Speech converts a voice but doesn’t translate. The Dubbing API does translate but creates the voice on its own, with no way to force a specific clone. Only Text-to-Speech could solve this: take the English .srt I already had, send each block to the TTS endpoint with my cloned voice, and assemble the audio aligned to the original video.

But having a translated .srt is not enough. Raw subtitle translation doesn’t survive TTS. Passages with source code, URLs, hex hashes, lists of shell commands — the model switches to spelling mode and the audio comes out at twice the expected duration. Translations that run too long blow past the time window and need to be condensed so the voice doesn’t sound like an auctioneer. Truncated SRTs need to be completed. And after each fix, you re-run the pipeline, listen, find the next problem, fix it, re-run. The whole thing was a cycle of interruptions, manual corrections, and re-runs — a far cry from the “press a button and get a dub” fantasy.

The big challenge was the accent. My cloned voice was trained on Brazilian Portuguese audio, so when it tries to speak English, the thick accent comes through. ElevenLabs has an [American accent] tag that works on v3, but on top of a voice trained in another language it’s weak — the Brazilian accent still bleeds through. The fix was to train a second voice of mine, English only. I recorded a few minutes in my best American accent, uploaded it as a separate Instant Voice Clone in my account, named it “Akita English”, and set the pipeline to use that voice by default. The result comes out more natural, no tag needed, and the voice identity is still mine.

The honest ceiling on my English voice quality

Let’s be frank. I’m a Brazilian who speaks English, not a native speaker, and I don’t have the time or patience to drill pronunciation for hours. A clone can only ever be as good as the source sample — if the source is a non-native reading a script, the clone inherits every imperfection. No clone is going to make me sound like “Fabio speaking perfect English” because that sound doesn’t exist for the model to learn from.

To get to the version that’s running today, I recorded five different training samples throughout the day, each with a different rhythm and script, and ran an A/B battery against the original Portuguese. None sounded native, and that was never the goal. The realistic bar was more modest: sound recognizably like me, read technical content without stumbling, and carry a full one-hour episode without wearing the listener out. The winner ended up being the first iteration, the original “Akita English”. Far from perfect, but good enough to ship. Swapping it out later is literally one line in the config file.

What surprised me most along the way was realizing that the rhythm of the training sample matters more than the timbre. The clone learns the cadence of whoever recorded it: iteration 5 came out 25% slower than iteration 2 reading the exact same text, purely because I recorded that particular sample more slowly. For dubbing tech videos, the right target is the rhythm of a “YouTube tech explainer” — brisk, energetic, around 150 words per minute. No audiobook pacing.

How I align audio to the original video

A non-obvious point: spoken English tends to be longer than equivalent Portuguese. If you just generate the audio for each subtitle and paste it at the original timestamp, the dub drifts later and later. With 30 minutes of video you’re already several seconds out of sync.

My approach was to attack the problem in layers. The first version split the SRT into smaller blocks, called “chunks”. Each chunk maxed out at 700 characters (so v3 wouldn’t hallucinate) and could only cut at a sentence boundary, never mid-sentence. A simple algorithm accumulated subtitles into a buffer until it was about to overflow, then walked back looking for the last cue whose text ended with ., !, or ?, and flushed only up to there. The cues still in the buffer got rolled into the next chunk. Each chunk also stored the start and end timestamps of the subtitles it covered, so the assembly step knew exactly where to drop that piece of audio later.

# For every incoming cue, check whether adding it
# would push the buffer past the char cap.
for cue in cues:
    projected = buffer_char_len(buf) + len(cue.text)
    if projected > max_chars and buf:
        # Walk back for the last cue in the buffer that
        # ends in '.', '!', '?', or '…'.
        sentence_end = last_sentence_end_index(buf)
        if sentence_end is not None:
            buf = flush(buf, sentence_end, chunks)
        else:
            # Run-on sentence longer than max_chars with no
            # terminator. Rare, but happens. Warn and cut.
            log.warning("run-on sentence > %d chars — "
                        "splitting mid-sentence as a last resort",
                        max_chars)
            buf = flush(buf, len(buf), chunks)
    buf.append(cue)

    # Soft-break: if the buffer is already at least 60% full
    # AND the current cue ended a sentence, flush now. Keeps
    # chunks balanced around 60-100% of max.
    if cue.ends_sentence and buffer_char_len(buf) >= max_chars * 0.6:
        buf = flush(buf, len(buf), chunks)

Technically, this was the turning point: instead of letting the dub follow the visual structure of the subtitle file, the chunker started following the real paragraph structure of the script. I won’t retell the whole story of how I got there and why it eventually forced a mass rerender, because I come back to that later in the section about the problems that only showed up after the first upload. The technical point here is simple: once the window represented a real spoken paragraph, the audio started sounding natural, and the stretch target could move from 92% to 95%.

Once the chunking is done, there’s still the problem that English runs longer than the Portuguese equivalent when spoken. The trick here is predicting whether a chunk is going to blow past its target window before generating the audio. If the ratio characters / 16 chars-per-second already exceeds the target duration by a margin, the pipeline passes the speed parameter directly to the ElevenLabs API, asking the model to natively generate faster speech. This preserves prosody way better than compressing afterwards. The ceiling is 1.15× (beyond that the voice starts sounding rushed).

# before calling the API, predict if it will overrun
expected_sec = char_count / EXPECTED_CHARS_PER_SEC  # 16 chars/s
predicted_ratio = expected_sec / target_sec

voice_speed = 1.0
if predicted_ratio > PREEMPTIVE_SPEED_THRESHOLD:   # 1.05
    voice_speed = min(predicted_ratio * 0.98,
                      PREEMPTIVE_SPEED_MAX)         # cap at 1.15
    voice_settings["speed"] = voice_speed

Even with preemptive speed turned on, sometimes the generated audio still comes out a tiny bit longer than the target window. That’s what the second safety net is for: measure the actual audio with ffprobe after it’s generated, compare to the target window, and apply the ffmpeg atempo filter to compress in post if it exceeded the 2% tolerance, capped at 1.20×. Combining native speed (1.15×) with atempo (1.20×) gives an effective compression ceiling of 1.38×, enough to fit naturally even in the worst cases without breaking quality.

# after generating the chunk
actual = ffprobe_duration(chunk_path)
ratio = actual / target_sec

if ratio > FIT_TOLERANCE:       # 1.02
    # compress the audio without touching pitch
    ffmpeg_atempo = min(ratio, FIT_MAX_ATEMPO)   # 1.20
    apply_atempo(chunk_path, ffmpeg_atempo)

When a generated chunk comes out shorter than the target window, the pipeline slightly stretches the audio (without affecting pitch) to reduce the ugly silence after it ends. The goal isn’t to fill the whole window — a natural pause between sentences is fine — just to smooth over the worst offenders that would feel like “cut audio”.

On a big episode, 95 minutes, 194 chunks, the total cumulative drift came out to -0.7%. About 43 seconds of drift across the entire episode, imperceptible while you’re watching.

What saved the budget on this architecture is that each chunk lives on disk as its own .mp3. The pipeline keeps a manifest with the normalized text of every chunk, and before calling the ElevenLabs API, it compares the current text against the cache. If the text hasn’t changed, it reuses the existing audio without spending a single credit. If I rewrite a problematic cue, only the chunks affected by that cue get regenerated — the rest of the episode stays untouched.

This is what made the iterations viable. I’d run the batch, listen to sections, spot a problem (translation too long, a code snippet the TTS couldn’t pronounce, a chunk boundary that landed at a bad spot), fix the SRT or tweak the chunker parameters, and re-run. Each re-run consumed a fraction of the credits and time of the original run, because only the changed chunks were regenerated. Without this cache, every iteration would have cost nearly as much as the first pass, and the total batch cost would have been two or three times higher.

That chunking change also had a heavy cache and cost impact, but I’ll leave that part for the later section where I explain the actual retracing and rerender cycle. Here, the important point is just that it was the right technical move.

When atempo can’t save you: rewriting cues before TTS

Not everything is about window alignment. There’s a kind of cue that breaks the whole pipeline, where neither preemptive speed nor post-process atempo can save it, and I only ran into it once I started running the batch over the more technical episodes of the channel.

The problem is this. The v3 reads at about 16 characters per second when you hand it normal conversational text. But when you hand it a cue stuffed with a literal URL, a hex hash, a binary string, a long sorted number list, or a shell code block, the model shifts into “spelling it out letter by letter” mode and drops to about 9 characters per second. A 500-character cue that was supposed to turn into 30 seconds of audio comes out at 55. The sanity check rejects it (because it’s past 1.8×), the automatic retry tries the same 500 characters again, the five retries all fail in a row, and the chunk gets stuck.

And this did not happen by accident. I used to write those passages that way because I knew the same scripts would later become blog posts, so it made sense to leave the raw URL, the full command, the complete hash, the whole technical detail there for the reader. In Portuguese that worked fine because it matched my original workflow exactly: record the video, then publish the matching text on the blog. What I had never prepared for was the next step, turning that same material into automated English dubbing. I only realized that conflict when the pipeline started failing for real.

I hit this first on ep052, the Ubuntu beginner’s guide, where two cues carried github.com URLs, hkp://keyserver.ubuntu.com URLs, and a 40-character GPG hash. Trying to fix that in post-processing is a waste of time. The 1.20× atempo ceiling will never compress 55 seconds into 30, and even if it did, the result would be an unlistenable chipmunk. The fix has to land upstream, in the text itself.

The solution was to rewrite the cue so it describes what the command does instead of showing the literal command:

- Run: apt-key adv --keyserver hkp://keyserver.ubuntu.com:80 \
-   --recv-keys 0A6A3E7F79F93EF8AAB9E92BAEBB74C8B5A1E44D
+ Run the full command in the video description to import the
+ signing key from the Ubuntu keyserver.

The on-screen English caption still shows the full command, so anyone reading the caption sees exactly what they need to type. The dubbed audio describes the intent in spoken language, which is what the TTS can deliver without choking. The instruction to the viewer stays intact, the characters-per-second ratio goes back to the expected 16, and the generated audio fits the original window without needing any stretch or compression in post.

I went hunting for cues like this across the whole batch. I scanned from ep052 through ep146 looking for cues with a high density of “hard” characters (digits, brackets, operators, long URLs, binary, hex) and ended up rewriting around 30 cues across 13 episodes:

• ep052 Ubuntu for devs: GitHub URLs and a 40-character GPG hash
• ep091 Hello World in C: binary examples
• ep095 640kB memory: address wrap, segment-offset
• ep106 CodeMiner: spelled-out email address
• ep113 Compression: pure binary strings, 75% hard characters
• ep115 SQL Server: long sorted number list
• ep120 Internet: dotted IP addresses
• ep121 Sockets: two JavaScript code blocks
• ep122 Proxies: Chrome User-Agent headers
• ep123 Secure networking: shell one-liners, Docker flags
• ep126 Gentoo: C chroot demo
• ep136 Containers: GitHub release URL
• ep144 Cryptography: signing key address

The pattern was always the same: keep the caption showing the code, URL, or hash verbatim, and rewrite the spoken text so the narrator describes what it does. I could have automated the rewrite by letting an LLM read each cue and propose a substitution on the fly, but it felt safer to review them all by hand. It’s exactly the kind of thing where the model decides to “improve” a command to make it cleaner and ships with broken shell. The manual sweep took about two hours. The automated version would have cost the same time in review, with the bonus anxiety of having shipped a mangled command by accident.

There was also one weird structural case on ep146, about Docker Compose. Two cues had glued themselves together because of a missing blank line in the SRT, and pysrt was treating them as one giant cue. The TTS never even got to process it — the chunker choked first. I fixed it by hand, adding the missing blank line. One-character fix, half an hour to track down.

Reconstructing truncated SRTs

Another trap showed up when I went to run the full batch: three episodes had English captions that were too short relative to the actual video length. Ep056, the Rails episode, was the most absurd case. Nineteen minutes of caption for an 80-minute video. Sixty-two minutes of content with no caption at all. Ep057 (WSL 2) was missing 14 minutes, and ep068 (Git Direito) was missing 2.

This is a leftover from my old translation workflow. At some point I started hand-revising the caption, stopped partway through, and the .en.srt file got saved truncated at the spot where I stopped. You can’t dub an 80-minute video with a 19-minute caption. The chunker produces audio up to where it can read and then simply has no idea what to do with the rest of the video.

The fix became a new script. It diffs the truncated .en.srt against the .pt-orig.srt (the raw YouTube auto-caption, which always covers the whole video), picks up the cues that only exist in Portuguese, sends them to Claude Sonnet 4.6 with a strict JSON schema asking for cue-by-cue translation, and pastes the result back at the tail of the English file. The schema is the detail that matters. When I tried asking for the response as plain text in N|text format, Claude was dropping about 20% of the cues along the way. With a strict JSON schema, the drop rate fell to zero across the three repairs.

The numbers:

• ep068 Git Direito: +50 cues / 2 minutes reconstructed
• ep057 WSL 2: +368 cues / 14 minutes
• ep056 Rails: +1445 cues / 62 minutes (plus dropping 5 fake cues the old translator had invented at the tail to plug the gap)

After the repairs, the three episodes joined the batch normally and got dubbed like any other. The kind of tool you hope you never need, but when you need it, it’s worth writing once and closing the problem for good.

Automatic emotion tags

While I was testing, I decided to add one more step to the pipeline: an “emotion tagger” that reads the English SRT before it goes to TTS and inserts tags like [sarcastic], [thoughtful], [emphatic], [deadpan] at spots where a human narrator would naturally shift tone. The idea is to mimic what a professional voice actor would do — hit the emphasis at the right moments — without turning the video into a theater of overblown emotions.

This part is tricky for two reasons. First: letting an LLM loose on your SRT runs a real risk of it “improving” the text (swapping a word here, rewriting a sentence there) and you ending up with a dub that doesn’t match the caption. Second: LLMs love to overdo it. Ask them to tag emotion and they’ll put a tag on every other line. To guard against both, I pinned a small allow-list of tags and run a round-trip validation after Claude’s response comes back:

ALLOWED_TAGS: frozenset[str] = frozenset({
    "[sarcastic]", "[thoughtful]", "[emphatic]",
    "[deadpan]", "[serious]", "[amused]",
    "[sighs]", "[exasperated]", "[confident]",
    "[matter-of-fact]",
})

def validate_tagged(original, tagged):
    """Ensure the tagged SRT preserves the original with no drift."""
    if len(original) != len(tagged):
        raise TagValidationError("cue count mismatch")

    for o, t in zip(original, tagged):
        # Index and timestamp must match byte for byte.
        if o.index != t.index or o.timestamp != t.timestamp:
            raise TagValidationError(f"cue {o.index}: header drift")

        # Any tag outside the allow-list invalidates the whole response.
        bad = find_disallowed_tags(t.text)
        if bad:
            raise TagValidationError(f"cue {o.index}: {bad}")

        # The text with the tags stripped must be byte-identical to
        # the original. If the LLM swapped a word, rewrote anything,
        # validation fails and the response is rejected.
        if strip_tags(o.text) != strip_tags(t.text):
            raise TagValidationError(f"cue {o.index}: text drift")

When validation fails, the pipeline throws the response away and re-requests (or, if it keeps failing, ships the SRT untagged). On tag density, the prompt sent to Claude explicitly says: aim for N/10 tags in a SRT with N cues, hard floor of 1 tag per 12 cues, hard ceiling of 1 tag per 7 cues, distribution balanced across the four quarters of the episode, and no two tagged cues within 4 cues of each other. That rule set was the sixth iteration — the five before it either blanketed everything in tags or loaded them all into the first quarter.

Don’t let v3 hallucinate

A critical detail that only shows up in production: ElevenLabs v3 tends to hallucinate when you send very long blocks of text. I’ve seen blocks where the input text was around 1,500 characters and the model generated nine minutes of audio when the expected duration was a minute and a half. The model just decides to keep talking on its own, inventing content.

The official ElevenLabs docs recommend keeping each call under 800 characters to avoid this. I went more conservative and cut at 700, always at sentence end (never mid-sentence, because a mid-sentence split sounds horrible when concatenated). On top of that, I bumped stability to 0.9 (Robust mode, more stable but less responsive to emotion tags), turned on apply_text_normalization so the model pronounces numbers and acronyms correctly, and added a sanity check that rejects any generated audio longer than 1.8× the expected duration.

DEFAULT_VOICE_SETTINGS = {
    "stability": 0.9,          # Robust mode
    "similarity_boost": 0.95,
    "style": 0.0,
    "use_speaker_boost": True,
}

EXPECTED_CHARS_PER_SEC = 16.0
CHUNK_SANITY_MAX_FACTOR = 1.8   # reject if actual > 1.8× expected

With those mitigations in place, a 95-minute episode (194 blocks) ran start to finish with zero hallucinations. Exactly one block failed mid-run on a transient 502 error from the API, which the automatic retry caught on the next attempt.

Mastering for YouTube

YouTube doesn’t publish an official number, but industry consensus is that its playback loudness normalization targets -14 LUFS. Deliver audio louder than that and YouTube turns it down; deliver it quieter and it leaves it alone. To land on the target exactly, the pipeline runs ffmpeg’s loudnorm filter in two passes. The first pass measures the entire audio and prints the stats (input_i, input_tp, input_lra, input_thresh) as a JSON blob on stderr. The second pass reads those stats back in and applies a linear static gain to land precisely at -14 LUFS, -1.5 dBTP:

# Pass 1: measure
ffmpeg -i final_en.mp3 \
  -af "highpass=f=80,loudnorm=I=-14:TP=-1.5:LRA=9:print_format=json" \
  -f null -

# Pass 2: apply linear gain with the measured values
ffmpeg -i final_en.mp3 \
  -af "highpass=f=80,loudnorm=I=-14:TP=-1.5:LRA=9\
:measured_I=-18.3:measured_TP=-3.1:measured_LRA=5.4\
:measured_thresh=-28.7:offset=1.2:linear=true" \
  -ar 48000 -ac 2 -c:a pcm_s16le final_en_mastered.wav

Two passes instead of one because single-pass mode runs in dynamic compression and ends up “pumping” the gain during silent stretches. With linear=true on pass 2, the gain stays static on top of the pass-1 measurements, so no pumping. The result lands within ±0.1 LU of the target, which is effectively inaudible. The highpass=f=80 filter in front kills HVAC rumble and mains hum below 80 Hz, which the human ear doesn’t catch but will move peak measurements around.

The point where I thought it was done. And it wasn’t.

This is where the most annoying, and most educational, part of the project starts.

I really thought I had nailed it on the first big pass. The pipeline closed all 146 episodes, I uploaded everything, and it looked done. The durations matched, the files were all there, and the whole process had finally crossed the complete batch of nearly 96 hours of video. But it was exactly after that upload that I started noticing problems I simply had not anticipated.

The first scare was the jingle situation. I already knew many Akitando episodes begin with that short instrumental intro before I start talking, but I underestimated how much trouble it would cause. YouTube’s .srt is almost useless here, because a jingle is not speech. Sometimes it marks [Music], sometimes it gives you nothing useful, sometimes the window is just wrong. In the first assembly pass, some episodes lost the jingle entirely, some got it doubled, and in others the source-fill from the original audio already contained the right jingle and my splice came in on top of it, pushing everything else forward. This was the kind of bug you only catch when you sit down and actually listen to the final mastered file, not when you just look at total duration.

The fix was to stop hoping the subtitle would tell me where the jingle was and go straight to the source. I built a detector using the original video audio itself plus reference jingles, with an audit and repair pass on top of the mastered outputs already generated. That eventually became a small subsystem of its own: detector, audit, cached reassembly, exact-window repair. At the end of that detour, the set of changed episodes collapsed into a well-defined reupload bucket and the jingle problem finally came under control. It was one of those classic production moments: the first solution looked good enough until it hit the whole archive.

Then came the second punch, and this one was more expensive: I had forgotten a basic fact about YouTube subtitles. SRT is built to be read on screen, not to become spoken script. So an idea that was a full paragraph in my original text showed up split into several small cues, each with its own micro time window. In the first version of the pipeline I was using those cues as the generation unit for ElevenLabs. The result was TTS audio that was coherent locally, but the final assembly inserted artificial silences in the middle of sentences. Technically aligned, humanly weird.

Fixing that hurt because it meant walking several steps back. Instead of respecting the cue structure of the SRT, I had to go back to my original scripts and rebuild chunking around the paragraphs I had actually read when recording the videos. Almost every Akitando episode has a matching blog post with a ## Script section, and that is what saved the project. Because I had been meticulous back when I made those videos — writing the scripts first, reading exactly that on camera, and archiving both the videos and the scripts — I had a source of truth. That made it possible to map the script paragraphs onto the timings in the .pt-BR.srt, regroup the .en.srt on top of that, and regenerate the chunks in the right structure.

The problem is that this destroys the previous cache. Even when the text barely changes, if the chunk boundary changes then the TTS call changes too: different context, different prosody, different beginning, different ending. So this was not a touch-up. It was a full regeneration of everything again. That was the point where the cost started escaping my original estimate.

And just when I thought now it really was over, the third hit landed: external clips. Many videos on the channel include inserted footage from other sources. The Ruby on Rails episode, for example, has 37signals clips, Apple material, commercials, external inserts in the middle of the narrative. YouTube’s .srt did not know what to do with that. In some places it simply hallucinated text, in others the timings were completely wrong, and in others the original speech had nothing to do with the subtitle I was using as ground truth. Again: if you only look at the automation, it looks fine; if you actually listen to the final master, it isn’t.

The fix, again, was to leave the subtitle behind and go back to the original material. I built an external-clip window detector based on alignment gaps between transcript, .pt-BR.srt, speechless cues and chunks that crossed those windows. Then, instead of trying to dub what made no sense to dub, the process started recovering the original audio from the source video exactly in those stretches and filling the master with the correct source audio there. That’s what fixed the 37signals, Apple Education, Mac vs PC commercial and other inserted segments I had forgotten were even there.

That was the biggest lesson from the second half of this project: automation takes you very far, but the full archive always has more dark corners than you imagine on the first pass.

The dubbing batch numbers (and why it hurt more than I expected)

I need to correct myself again. I didn’t just miscalculate the first bill. I also overestimated how confident I should have been after the first full run.

When I upgraded to ElevenLabs Business, I thought I would finally have plenty of headroom. In practice, I didn’t. The latest dashboard screenshot shows 10,046,279 credits consumed out of a limit of 11,000,000, leaving 953,721:

In plain English: I burned through more than 91% of the Business workspace limit. And that helps explain what “10 million credits” really means in this context. It does not mean “I rendered 146 episodes once”. It means rendering 146 episodes, discovering a structural problem, going back, fixing it, regenerating chunks, discovering another structural problem, going back again, rebuilding masters, auditing jingles, repairing external clip windows, uploading again. Every extra round costs money. And when the change invalidates the cache, it costs a lot.

The cost also doesn’t stop at ElevenLabs. All the subtitle curation, problematic cue rewrites, truncated SRT reconstruction, jingle audits, external clip detector work and the rest ran through Claude Code on the Claude Max 20× plan. Combined with the mass blog translation to English that I described in the anniversary post, the Claude Max 20× weekly limit hit 100%, with more than BRL 300 in extra usage on top. If you think generative AI becomes free once you’ve signed up for the plan, the invoices disagree.

Even so, I still think it’s worth it. We’re talking about 146 episodes, almost 96 hours of technical content, more than 5 years of channel archive. Doing that manually in a studio, with an actor, direction, pickups and human review on everything, would cost a completely different order of magnitude.

What changed was how I think about the project. At the start I was looking at this as “one very large batch”. In the end it was something else. It was digging through a whole archive, finding hidden exceptions, fixing old rot, rebuilding masters, rerendering, relistening.

The first full batch took a little under two days and made me think I had solved everything. I hadn’t. The second pass, with paragraph-aware chunking, was already expensive. The third, with jingle repairs and mastered-output fixes, cost more again. And the fourth hit came from the external clips I had not modeled correctly in the first automation pass.

But now, finally, I think it’s done. Or rather: I hope it’s done. Everything has been uploaded again, for the last time. It isn’t perfect in the absolute sense of the word. It’s as good as I can get it with scripts, audits and automation, without descending into manual-intervention hell episode by episode.

And if all of this was possible, it’s because years ago I had the discipline to produce the videos the right way: write first, read the script when recording, and archive the materials afterward. Without the original videos and without the original scripts, I would have had no way to discover where the .srt was lying, where the timing was wrong, where the external clip started, where the original paragraph began and ended. That old organization is what made it possible to repair the project now.

So yes: the dub for all 146 episodes is done. Not perfect. But closed. At least, I sincerely hope this time it really is.

The first test, watch it

I made a test video to prove the thing works. Two caveats up front: first, the embed below already loads with English captions on by default, so that side is solvable via URL. Second, the audio is annoying: YouTube removed the audio track selector from embedded players back around March of this year. So inside the embed you won’t find that option in the gear menu at all, it just plays in Portuguese. To actually hear the English dub, click watch directly on YouTube — the main YouTube watch page still has the audio track switcher in the gear menu.

If you’re used to YouTube’s automatic dubbing or to the AI dubbing on TikTok, compare them. The difference is glaring. The voice is mine, with a worked-over American accent, and the sync with the original video stays within 1 to 2% cumulative drift, practically imperceptible over 95 minutes of video.

And the missing piece: translating the thumbnails

While I was closing this post, the penny dropped: I’d left a loose end. All 146 thumbnails on my channel are in Portuguese, many of them with big all-caps titles like “9 DICAS PARA PALESTRANTES” or “7 RECOMENDAÇÕES DE SHOWS PARA PESSOAS DE TECH”. Perfect English audio doesn’t help if the image in the search results is an illegible block of text to anyone who doesn’t read Portuguese. YouTube lets you upload an alternative thumbnail per language, so I needed a way to generate English versions, keeping the rest of the art identical and swapping only the text.

The tool uses two pieces. yt-dlp is the modern fork of the old youtube-dl, a Python CLI that downloads anything from YouTube (videos, audio, subtitles, thumbnails) without needing an API key. Nano Banana Pro (nano-banana-pro-preview on the API) is Google Gemini’s latest image-editing model — you feed it an input image with a prompt and it returns an edited image, preserving the rest of the composition when you ask it to touch only a specific part.

The pipeline has two steps. Step 1: yt-dlp grabs the thumbnail as a .jpg, no video download:

yt-dlp --skip-download \
       --write-thumbnail \
       --convert-thumbnails jpg \
       -o "thumbnails/originals/<slug>/<video_id>" \
       "https://www.youtube.com/watch?v=<video_id>"

Step 2: send each image to Gemini Nano Banana Pro with a prompt that’s a rigid contract of hard requirements, not just a “translate this”:

TASK:
Detect every piece of Portuguese text visible on this image and
translate it to clear, natural American English. Replace the
Portuguese text with the English translation in the same visual
position.

STRICT REQUIREMENTS — follow every one:
- Preserve EVERY non-text element identically: face, pose,
  expression, background, color palette, lighting, icons, logos,
  decorative shapes, borders, layout. Only the text changes.
- The English translation must be IDIOMATIC and CONFIDENT — not
  a literal word-for-word rewrite. It's a YouTube thumbnail for a
  tech audience, so use punchy phrasing a native English-speaking
  tech YouTuber would write.
- Match the original text's font family, weight, size, color,
  stroke outline, drop shadow, and any decorative treatment.
- If there is no Portuguese text at all on the image, return the
  original image unchanged.

The “return the original image unchanged” clause is what saves the earliest episodes, which don’t have overlay text, just my face. Without it, the model would invariably try to “improve” the composition, swap the lighting, and so on.

Here’s the result on two examples (episodes 10 and 11):

Original (PT)

Translated (EN)

Original (PT)

Translated (EN)

Notice the detail on episode 10. The model didn’t translate “7 Recomendações de Shows para Pessoas de Tech” literally into “7 Show Recommendations for Tech People” the way Google Translate would. It rewrote it as “7 TV Shows You Must Watch If You’re In Tech”, the way an actual American tech YouTuber would headline the video. The rest of the image stays pixel-for-pixel identical. On episode 11, “9 DICAS PARA PALESTRANTES: venda sua caneta” becomes “9 TIPS FOR SPEAKERS: sell your pen”, with the font, all-caps treatment and position preserved. If you didn’t know the original was in Portuguese, you couldn’t guess the second image is an automatic AI edit.

Cost on the Gemini side: a few cents per image, a handful of dollars total for the whole batch. Trivial next to the $1,500+ on the audio dub. With the thumbnail sorted, the channel’s English conversion is finally buttoned up — ElevenLabs v3 cloning the voice over hand-curated English .srt files, and Nano Banana Pro editing the image so the title matches the audio.

The conclusion, which is the title of this post

When Qwen3 TTS dropped, the hype was calling it an “ElevenLabs killer”. I spent weeks trying to live with that premise in practice, with real content shipping every week. And what I found is that open source TTS is still miles behind ElevenLabs. The gap is large, and it’s a gap that shows up precisely when you step off the 5-second demo tweet and put the model to work on a real 30-minute weekly podcast.

In practice, Qwen3 doesn’t even beat ElevenLabs’ v2 model. v3, which is the current one, is still a step above v2. The prosody comes out better, the inline emotion tags work in Portuguese and English without effort, and the API stays up without you maintaining a server. The cost per character is a bit higher, but for my current volume it sits comfortably within budget and gives me back the hours I was spending on GPU babysitting.

The lesson here is the same one the LLM crowd is learning slowly. A 30-second demo tweet is one thing, production is a completely different thing. Open source has its niche, mainly when you have sensitive data that can’t leave the house, or when you have tons of idle GPU and little recurring budget. But for serious TTS use in a commercial product, here in April 2026, ElevenLabs is still untouchable. Qwen3 didn’t kill anyone.

And don’t forget: subscribe to The M.Akita Chronicles on Spotify so you don’t miss new episodes like this one, made with the new pipeline.

⚠️ Obsolete article (updated 2026-04-24). The conclusions and rankings in this post were superseded after I re-audited the benchmark against the ruby_llm gem source and restructured the evaluation criteria. Several “hallucinations” I had cataloged were actually valid API. Kimi K2.6 and Gemini 3.1 Pro moved up to Tier A. GLM 5.1 dropped to Tier C. MiMo V2.5 Pro fell from “first non-Anthropic Tier 1” to Tier B. The canonical version lives at LLM Coding Benchmark (April 2026). This post stays as a historical record of what I concluded before the re-audit.

Update April 16, 2026: Added Claude Opus 4.7, Qwen 3.6 and GPT 5.4 via Codex CLI (xHigh reasoning). Opus 4.7 is an incremental improvement over 4.6 (28 tests vs 16, same correct API) and becomes the new baseline. GPT 5.4, previously Tier 1 based on my personal vouch, now has objective data and dropped to Tier 2 — burned 7.6M tokens (~$16/run, 15x more expensive than Opus) and got the add_message calling convention wrong on multi-turn. Qwen 3.6 Plus remains Tier 3 with the same API hallucination as 3.5. The conclusion stands: if you want safety, pick Opus.

TL;DR: If you don’t want to read the whole analysis: the only models that produced code that actually works in our benchmark were Claude Opus 4.7, Claude Opus 4.6, GLM 5 and GLM 5.1 (from Z.AI, ~89% cheaper than Opus). Sonnet works in this benchmark too, but in practice it falls short on projects requiring deeper reasoning — details in the conclusion. GPT 5.4, previously Tier 1 based on my personal vouch, now has objective data via Codex CLI: it burned 7.6M tokens ($16/run) and got the add_message calling convention wrong — works on the first message, breaks on multi-turn. Dropped to Tier 2. Everything else — Kimi, DeepSeek, MiniMax, Qwen, Gemini, Grok 4.20 — invented APIs that don’t exist or ignored the gem we asked for.

There’s a new wrinkle in this update: I redid the local part of the benchmark on an RTX 5090 (instead of the AMD Strix Halo) and added a fresh batch of Qwen models, including a Qwen 3.5 27B distilled directly from Claude 4.6 Opus. That reopened the conversation on running open source models locally. The 5090’s memory bandwidth flips the game from “unworkable” to “workable with 1-2 follow-up prompts.” The bottleneck for open source models has moved to a lack of factual knowledge about specific libraries, which I unpack in detail in the new section on the Qwen family. The Claude distillation gamble, by the way, gave a pretty frustrating result that I haven’t seen documented in these terms before.

If you’ve been following my previous vibe coding pieces, you know I spent the last two months in a 500-hour marathon using Claude Opus as my main coding agent. The results were good, as I reported in the conclusion about business models. But there was an itch I couldn’t scratch: am I locked into one model? Is there a real alternative to Claude Opus for daily use on real projects?

I’ve got an RTX 5090 with 32 GB of GDDR7. I know I can run the latest open source models. I bought a Minisforum MS-S1 with an AMD Ryzen AI Max 395 and 128 GB of unified memory, and built a home server with Docker to serve local models. The infrastructure was ready. What was missing was actually testing it.

I built an automated benchmark to compare open source and commercial models under identical conditions. 33 models configured in total (25 from the original run plus 8 added in the NVIDIA rerun), 27 executed, 16 completed in some form. The code is on GitHub.

The bottleneck nobody explains: VRAM and KV Cache

Before getting to the results, I have to explain why running large models locally is much harder than it looks.

Take Qwen3 32B. The model in FP16 (full precision) takes ~64 GB. Quantized to Q4 (4 bits), it drops to ~19 GB. So it fits in my RTX 5090’s 32 GB, right? Wrong. That’s just the model weights. There’s a part nobody tells you about: the KV Cache.

KV Cache is the memory the model uses to “remember” what it has already read. Every time it processes a token (a word or piece of a word), it computes two vectors — K (key) and V (value) — for every attention layer. Those vectors stick around so the model doesn’t have to recompute everything when it generates the next token. Without that, generation would be quadratically slow.

The KV Cache scales linearly with the size of the context. The formula:

KV Memory = 2 × Layers × KV_Heads × Head_Dimension × Bytes_per_Element × Context_Tokens

For a model like Llama 3.1 70B in BF16, that comes out to ~0.31 MB per token. Sounds tiny, until you realize that a 128K context eats 40 GB of KV Cache alone. The model itself plus KV Cache adds up to way more VRAM than most GPUs have.

And for actual coding agent use, 128K tokens isn’t a luxury, it’s the bare minimum. The agent has to read files, keep conversation history, receive command output. In long benchmark sessions, our models consumed between 39K and 156K tokens. Less than 100K of context isn’t practical for day-to-day project work.

Google published TurboQuant (ICLR 2026), which compresses the KV Cache to 3 bits without accuracy loss — a 6x memory reduction and up to 8x speedup. It uses random vector rotation (PolarQuant) followed by a 1-bit algorithm on the residuals. Works online during inference, compressing on write and decompressing on read. Not yet implemented in the runtimes we use (llama.cpp, Ollama), but when it lands it’ll change the equation a lot.

For anyone wanting to dig deeper into the VRAM math, I recommend this link from Ahmad Osman for the article “GPU Memory Math for LLMs (2026 Edition)”.

The hardware problem: not all memory is created equal

“But I have 128 GB of RAM!” Cool, but that’s not what matters. What matters is memory bandwidth, and the difference between types is wild:

The RTX 5090 has 7x the bandwidth of the LPDDR5x memory in my Minisforum. That means even if a model fits in the AMD’s unified RAM, inference will be proportionally slower. On my Minisforum with LPDDR5x at 256 GB/s, Qwen3 32B runs at ~7 tok/s. On the RTX 5090 at 1,792 GB/s, it’d be much faster — if it fit entirely in VRAM alongside the KV Cache.

Most folks running local models are still on DDR4. At 50 GB/s, 32B models are basically unusable. And there’s another factor people forget: storage. When the RAM can’t keep up and the system swaps, the storage speed becomes the bottleneck:

From SATA to NVMe Gen5 you’re looking at a 22x difference. If you’re doing partial offloading to disk (which is common when the model doesn’t fit entirely on the GPU), NVMe Gen4 or Gen5 makes a real difference. SATA is a non-starter.

To sum up: running local models isn’t just “having enough RAM.” You need the right kind of memory, with the right bandwidth, and fast storage as a fallback. For a lot of people, a Mac Studio with high-bandwidth unified memory (up to 800 GB/s on the M4 Ultra with 512 GB) would be the more practical option, but it costs more than US$ 10,000. The AMD Ryzen AI Max is the cheaper alternative with unified memory, but its LPDDR5 caps out at 256 GB/s.

Ollama vs llama.cpp: why Ollama falls apart on benchmarks

Ollama is the most popular way to run local models. Install, pull the model, run. For casual use it works. But when I tried to use it for automated benchmarks with long unattended sessions, it broke in 6 different ways across 8 models:

1. Unloads the model mid-session. On long runs, Ollama decides the model isn’t being used and unloads it from the GPU. The agent sits there waiting for a response from a model that no longer exists.
2. Ignores the requested context. You ask for num_ctx=131072, Ollama accepts, then halfway through the run it reverts to the default without warning.
3. Unstable lifecycle. Asking for keep_alive: 0 to unload doesn’t always work. The model stays resident and blocks the next one.
4. Incompatible formats. Native bf16 variants on Ollama failed, while the same model as a Q8 GGUF from HuggingFace worked fine.

The fix: migrate to llama-swap, a Go wrapper that manages llama.cpp processes with hot-swap. A request comes in for a different model than the one currently loaded, it kills the current process and starts the new one. No context negotiation, no flaky lifecycle.

llama-swap fixed the loading of 6 of the 8 models that had failed under Ollama:

But llama-swap isn’t magic.

Why “just use llama.cpp” doesn’t fix everything

llama.cpp solves Ollama’s lifecycle problems but brings its own:

Each model needs specific flags. GLM and Qwen 3.5 emit <think> tags that break clients if you don’t pass --reasoning-format none. Gemma 4 needs build b8665+ for the tool call parser to work.

Not every model supports tool calling. llama.cpp needs a dedicated parser for each model’s tool call format. Llama 4 Scout uses a “pythonic” format ([func(param="value")]) that llama.cpp simply doesn’t parse and emits as plain text. vLLM has a parser for it, llama.cpp doesn’t.

And then there are the repetition loops. Gemma 4, even with the right parser, gets into an infinite loop after ~11 tool calls in long sessions. It’s a known bug that PR #21418 didn’t fully fix.

Tool calling compatibility per model:

At the end of the day, llama.cpp is better than Ollama for automated runs, but “plug and play” it ain’t. Each model requires specific configuration, and some just don’t work for agentic coding yet.

Reasoning: models that think vs models that wing it

There’s one difference between models worth explaining: reasoning. The idea is that the model “thinks before answering” instead of generating tokens straight from left to right. Models with reasoning go through an internal chain-of-thought step where they evaluate the problem, consider alternatives, plan, and only then emit the response.

In practice this shows up as <think>...</think> tags in the output, blocks of text the model writes to itself that shouldn’t go to the end user. Claude Opus 4.6, GPT 5.4, DeepSeek V3.2 and the Qwen 3.5 line support reasoning natively. The smaller ones (Gemma 4, GPT OSS 20B, older models) don’t have that capability.

Why does it matter for coding? When a coding agent gets “build a Rails app with 9 components,” it has to decompose the task into steps, decide the order, anticipate dependencies, adapt when something fails. Without reasoning, the model generates code sequentially with no planning. It works for simple tasks, falls apart on projects with interdependent parts.

In the benchmark, the difference was clear:

• GPT OSS 20B (no reasoning, 20B parameters) created the app in the wrong directory. Couldn’t keep workspace instructions in mind while generating code.
• Qwen 3 32B has reasoning, but at 7 tok/s it was too slow. The “thinking” tokens drag out the generation time.
• Gemma 4 31B, with no reasoning trained for agentic use, fell into repetitive tool calling loops.
• GLM 5 (cloud, 745B MoE) with reasoning and 44B active parameters, finished cleanly and used the correct API.

There’s a trade-off: reasoning consumes extra tokens (the <think> blocks), which take up VRAM in the KV Cache and slow generation down. That’s why flags like --reasoning-format none are needed in llama.cpp. Some clients don’t know what to do with reasoning tokens and break. Models that emit reasoning when the runtime isn’t expecting it can produce garbage in the output.

And reasoning isn’t something you “turn on” in any model. It’s a capability trained with reinforcement learning on top of the base model, using data from problems that require multi-step thinking. The smaller open source models (20B-35B) typically didn’t go through that training, or went through it on a smaller scale. They know how to generate code, but they don’t know how to plan code. On tasks that require 50+ coordinated tool calls, that difference is fatal.

The benchmark: methodology

To compare models fairly, I built an automated harness in Python. Each model gets the exact same prompt: build a complete Ruby on Rails application, a ChatGPT-style chat SPA using the RubyLLM gem, with Hotwire/Stimulus/Turbo Streams, Tailwind CSS, Minitest tests, CI tools (Brakeman, RuboCop, SimpleCov, bundle-audit), Dockerfile, docker-compose and README.

The runner is opencode, one of the most popular open source coding CLIs, competing with Claude Code and Codex. One clarification worth making: the original author left the project to work on Crush together with the Charm team (the folks behind Bubble Tea, Lip Gloss and several other Go terminal tools), but the rest of the original team kept evolving opencode normally — the project has not been discontinued. Today the two coexist. If you read my piece on Crush, you already know that branch. Both run everywhere: macOS, Linux, Windows, Android, FreeBSD.

I actually tried to use Crush for the benchmark first. The problem: it advertised a --yolo flag in its help to auto-approve every action (essential for unattended automated runs), but at runtime it rejected the flag. Without auto-approve there’s no way to do an unattended benchmark. opencode, on the other hand, had the opencode run --agent build --format json mode that emits JSON events with session IDs and token counts, perfect for automation. So we went with opencode.

I picked opencode (and not Claude Code or Codex) for two reasons:

1. Neutrality. Claude Code is optimized for Anthropic models. Codex is optimized for OpenAI models. opencode is agnostic, same interface for all.
2. Automation. opencode exposes a machine-readable JSON format. Claude Code and Codex don’t have an equivalent interface for external benchmarking.

Cloud models ran in two phases: phase 1 (build the app) and phase 2 (validate local boot, docker build, docker compose). Local models only ran phase 1.

Worth mentioning: the entire benchmark cost less than $10 in tokens on OpenRouter. Apart from GPT 5.4 Pro which torched $7.20 to fail, the other 11 cloud models added up to about $2.50 total. Local models cost only electricity. The point is: running your own benchmark is cheap. If you want to know whether a model works for your use case, drop the $2 and test it. The harness code is on GitHub, just swap the prompt for your own project.

Why GPT 5.4 failed the benchmark (but not in real life)

GPT 5.4 Pro is the only cloud model that consistently failed our benchmark. Two separate runs, same result: the model generated files but never reached finish_reason: stop. It always ended on finish_reason: tool-calls — wanted to keep calling tools but the loop kept breaking.

For folks who don’t know: tool calling is when an LLM needs to perform an action (read a file, run a command, edit code) and emits a “tool call” in a structured format. The client (opencode, Claude Code, Codex) interprets it, executes it, and returns the result back to the model. Each provider has its own format: Anthropic uses tool_use blocks, OpenAI uses function_calling with proprietary JSON schemas, Google uses FunctionCall.

GPT 5.4 is heavily trained for OpenAI’s native function calling format — tool_choice, tools with proprietary JSON schemas. When the benchmark routes through opencode → OpenRouter → GPT 5.4, the tool schemas get translated at every hop. If GPT emits tool calls in a format that OpenRouter or opencode doesn’t parse correctly, the agent loop breaks.

The evidence: every other cloud model (Claude Opus, Claude Sonnet, Kimi K2.5, DeepSeek V3.2, MiniMax M2.7, GLM 5, Qwen 3.6 Plus, Step 3.5 Flash) ended on finish_reason: stop. Only GPT ends on finish_reason: tool-calls.

A fair comparison for GPT 5.4 would require running it in its native environment. And now we have that comparison: we built support to automate the Codex CLI (codex exec with --dangerously-bypass-approvals-and-sandbox and reasoning effort xhigh) and ran the same benchmark. GPT 5.4 completed in 22 minutes, generated all 9 artifacts, wrote 22 tests with the most sophisticated architecture in the entire benchmark: dependency injection of the RubyLLM client, PORO models for ChatMessage and PromptSubmission, session-backed ChatSession with TTL and message trimming, bin/ci script.

But the code breaks on the second message. GPT 5.4 uses chat.add_message(role:, content:) with keyword arguments instead of a positional hash chat.add_message({role:, content:}) — this causes ArgumentError: wrong number of arguments (given 0, expected 1) on the first multi-turn exchange. The first message works (uses chat.ask directly), multi-turn doesn’t.

And the cost: 7.6 million tokens on xHigh reasoning effort. That’s 65x more tokens than Opus 4.7 (118K) for the same benchmark. Estimated cost of ~$16 per run, versus ~$1.10 for Opus. Spent 15x more and still got the calling convention wrong. A massive token budget and maximum reasoning effort still can’t guarantee factual correctness on a gem API. API knowledge is binary recall in the weights, not a function of how hard the model “thinks.”

With objective data in hand, GPT 5.4 moves from Tier 1 to Tier 2. The architecture it generates is better than Opus in terms of design patterns. But the code needs a fix for multi-turn to work, and the per-token cost is prohibitive.

Sonnet and Opus through opencode/OpenRouter were probably also not pushed to their limits. Claude Code offers native tool support that opencode doesn’t replicate — meaning the benchmark results represent a floor, not a ceiling, for those models.

Open source models: reality vs the narrative

A lot of people are saying open source models have already caught up with the commercial ones and you can run your own “Claude” at home. In practice, not really.

The scale isn’t comparable. Frontier models like Claude Opus 4.6 and GPT 5.4 are closed-source, but estimates put them in the hundreds of billions to trillions of parameters range, trained with compute and data no open source company can replicate. The best models that fit on reasonable hardware are:

Post-publication correction: Qwen 3.5 35B is actually the 35B-A3B, an MoE with only 3B active parameters per token (not dense, as I’d originally written). That’s why it runs relatively fast for its size. And for folks with 24 GB of VRAM, the model recommended by Unsloth themselves is the Qwen 3.5 27B dense — that one I didn’t get around to testing in the benchmark, but it’s worth a look. For anyone wanting to dig deeper into local models, @sudoingX has been doing some serious experimentation in this space. Thanks to @thpmacedo for the heads-up.

Even the largest open source MoE (Mixture of Experts) models that companies make publicly available activate only a small fraction of parameters per token:

These large models aren’t self-hostable. Kimi K2.5 with 1T parameters needs GPU clusters with hundreds of GBs of VRAM. GLM 5 with 745B is the same. Even if Alibaba or Z.AI release the weights (and some do), nobody has home hardware to run them.

What fits on your home GPU are the 20B-35B models — and those have real limitations.

What each local model did in the benchmark

Results from the original run on the AMD Strix Halo:

Qwen 3 Coder Next (30B) — Completed in 17 minutes on the Strix, generated 1675 files, Rails app with all the artifacts. But only 3 tests. And more importantly: it invented RubyLLM::Client.new, a class that doesn’t exist in the gem. The app doesn’t run.

Qwen 3.5 35B — Completed in 28 minutes on the Strix, 1478 files, 11 tests. Used RubyLLM.chat without a model parameter — works only if the default is configured. No LLM mocking in the tests.

Qwen 3.5 122B — Completed in 43 minutes on the Strix, 1503 files, 16 tests. But it ignored the RubyLLM gem completely and built a custom HTTP client for OpenRouter. The prompt explicitly asked for ruby_llm.

GLM 4.7 Flash (local, Strix) — Produced 2029 files with all the artifacts, but the session ended mid-tool-call. The cloud version (GLM 5) works perfectly.

Gemma 4 31B (Strix) — Infinite tool call loop after ~11 productive steps. Known llama.cpp bug.

GPT OSS 20B (Strix) — Created the Rails app in the wrong directory (project/app/ instead of project/). A 20B model doesn’t follow workspace instructions reliably.

Qwen 3 32B (Strix) — Way too slow (7.3 tok/s). The hardware can’t keep up.

And the results from the rerun on the NVIDIA RTX 5090 (all with Q3_K_M or Q4_K_M and contexts between 64k and 128k to fit the 32 GB of VRAM):

Qwen 3.5 35B-A3B (5090) — 5 minutes at 273 tok/s. Recognizable Rails project, entry point RubyLLM.chat(model:) is right, but it hallucinates chat.add_message(role:, content:) and chat.complete instead of .ask. Fixable in 1-2 follow-ups. The best candidate for “OSS local that’s actually worth trying.”

Qwen 3.5 27B Claude-distilled (5090) — 12 minutes at 129 tok/s. Impeccable Claude style, total API hallucination (RubyLLM::Chat.new.with_model{}, add_message, response.text). More details in the distillation section below.

Qwen 3 Coder 30B (5090) — 6 minutes at 145 tok/s. Returned a hardcoded mock string instead of calling the API. Tier 3 unusable.

Qwen 2.5 Coder 32B (5090) — 90 minutes of timeout, zero files. The model spun without ever calling a write tool.

Qwen 3 32B (5090) — 4 minutes at 69 tok/s, partial scaffold, errors. The general version is better than the Coder one but still breaks.

Gemma 4 31B (5090) — 8 minutes at 213 tok/s. Same repetition loop it had on the Strix. The llama.cpp bug isn’t a hardware issue.

Qwen 3.5 27B Sushi Coder RL (5090) — Infrastructure failure (ProviderModelNotFoundError), couldn’t be evaluated. Redo on a future run.

GPT OSS 20B (5090) — Pulled from this run because of a recent llama.cpp main regression in the harmony family tool call parser. The logs show Failed to parse input at pos 755: <|channel|>... in multi-turn sessions. It worked on the Strix with llama.cpp b8643, broken on today’s main. Waiting on upstream to fix it.

Cloud models: what actually works

Of the 12 models that completed the benchmark, all of them generated a recognizable Rails project with all the requested artifacts (Gemfile, routes, views, JS, tests, README, Dockerfile, docker-compose). 9 out of 9 on the completeness checklist.

But here comes the question that matters: does the code run?

The correct RubyLLM API is simple:

chat = RubyLLM.chat(model: "anthropic/claude-sonnet-4")
response = chat.ask("Hello")
response.content  # => "Hi there!"

8 of the 12 models invented APIs that don’t exist. The most common pattern: hallucinating an interface that doesn’t match the actual gem:

The models that got it right — both Claudes, GLM 5 and GLM 5.1 — used the simple two-step pattern (chat = RubyLLM.chat(model:) then chat.ask(message)). The ones that got it wrong tried to make RubyLLM look like the OpenAI Python SDK, which is a different thing. Grok 4.20 was the most brazen case: it didn’t even try to use the gem, it went straight for OpenAI::Client pointing at the OpenRouter URL, ignoring the explicit prompt.

And the tests? Only Opus, Sonnet, GLM 5 and GLM 5.1 did proper mocking of the LLM calls. All the others either hit the real API (which fails without a key) or mocked the invented API (tests pass but prove nothing). Test count is a misleading metric: Kimi K2.5 wrote 37 tests, more than anyone else, but none of them test real functionality because the API it uses doesn’t exist.

Real viability table

*Step 3.5 Flash works by calling the OpenRouter REST API directly with Net::HTTP, completely bypassing the gem the prompt asked for.

Now, this doesn’t mean those models are useless. If you take Kimi K2.5 or DeepSeek V3.2 and tell it “the RubyLLM::Client class doesn’t exist, fix it to use the gem’s real API”, it’ll probably fix it. One or two follow-ups and the project becomes functional. Most of the models that failed here could deliver a working project with a few more rounds of conversation.

But that’s where the trade-off lives. With Opus or GPT 5.4, the first output already works. You ask, they deliver, you test, it runs. With the cheaper models, you’ll spend time fixing API hallucinations, debugging code that “looks right” but crashes, steering the model in the right direction. Each of those rounds is 10-30 minutes. Three extra rounds and you’ve spent an hour of your time to save $0.90 in tokens.

You save dollars, you spend time. And time is money. For someone learning or exploring without urgency, that trade can make sense. For someone who needs to ship, the frontier models pay for themselves fast.

Comparing the models that work

Full ranking by time and tokens

DeepSeek V3.2 is the slowest despite being cloud — it has no prompt caching, so it resends the full context on every turn.

Token efficiency and cache

Models with prompt caching pay much less in effective tokens:

Speed: the chasm between cloud and local

There’s an aspect that the cost tables hide: inference speed. And the difference is brutal.

Claude Sonnet generates 532 tok/s. Qwen 3.5 122B running locally on my Minisforum (AMD Strix Halo) generates 22 tok/s. That’s a 24x difference. In practice, what Sonnet does in 16 minutes, Qwen 3.5 122B takes 43 minutes. Qwen 3 Coder Next at 37 tok/s is the fastest of the local models on the Strix and even so it’s 14x slower than Sonnet.

And it’s not just clock time. When you’re in an interactive coding loop — ask for a change, wait for output, test, ask for another — the model’s speed sets your rhythm. At 37 tok/s, every long response makes you wait 30-60 seconds. At 530 tok/s, it appears almost instantly. Over a day, you feel it.

DeepSeek V3.2 is a curious case: it’s cloud but it runs at 53 tok/s, slower than the locally-running Qwen 3.5 35B on the Strix (46 tok/s). The reason is that DeepSeek has no prompt caching — it resends the full context on every turn, strangling throughput. Paying for a cloud model that’s slower than running it locally doesn’t make any sense.

Local models are free in tokens, but they pay in time. On the AMD Strix, that math was a non-starter for every Qwen I tested: two minutes waiting for a long response, multiplied by 50 turns, eats your whole afternoon. But that changes when the hardware changes, and that’s why I redid the local part of the benchmark on a different machine.

AMD Strix Halo vs NVIDIA RTX 5090: what changes when the memory bandwidth doubles

To check whether the bottleneck was hardware or model, I took the same Qwen models and reran the benchmark on a workstation with an NVIDIA RTX 5090 (Blackwell, 32 GB GDDR7, 1,792 GB/s bandwidth). The numbers shift in a way that’s worth looking at carefully.

To put those speeds in context, remember that in the cloud Sonnet runs at 532 tok/s, Opus at 347 tok/s, Step 3.5 Flash at 242 tok/s, Gemini 3.1 Pro at 128 tok/s and Kimi K2.5 at 160 tok/s. Qwen 3.5 35B-A3B on the 5090, at 273 tok/s, is in the same neighborhood as Step 3.5 Flash, faster than Gemini, Kimi and GLM 5.1. Qwen 3 Coder 30B at 145 tok/s is in Gemini territory. The classic line “local models are ten times slower than cloud” stopped being true the moment the 5090 entered the conversation.

The practical consequence is that the “time is money” argument shifts. On the Strix, “waiting an hour for a Qwen 3.5 122B to do what Sonnet does in 16 minutes” is straight-up loss. On the 5090, waiting 5 minutes for Qwen 3.5 35B-A3B to do the work, plus 10-15 minutes for you to do 1-2 correction prompts, gives you a total in the 20-25 minute range. Sonnet does it in 16 minutes with zero corrections. The difference shrank enough that, if cost matters a lot, it’s worth it.

The catch: for this to be worth it, the model has to be close enough to the right answer that 1-2 correction prompts can fix it. When the error is “the model decided not to use the gem I asked for and returned a hardcoded mock string,” like Qwen 3 Coder 30B did, no easy correction prompt fixes that. That’s a redo.

Before you spend money on hardware thinking it’s the answer

I’ve got to give a warning here, because it’s the most common buying mistake I see right now. Every other week somebody tells me they’re going to grab a Ryzen AI Max because it has 128 GB of unified memory and that “lets you run huge models.” Technically, sure — the model fits. In practice, it’s almost unusable. The memory is LPDDR5x at 256 GB/s, seven times slower than the 5090’s GDDR7. What fits doesn’t run at human speed. My own Strix with Qwen 3.5 122B hit 22 tok/s and the run took 43 minutes. To do anything serious day to day, that’s not workable.

The 5090 is clearly superior, and it starts to make sense even for smaller models precisely because of the memory bandwidth. A Mac Studio with high-speed unified memory (up to 800 GB/s on the M4 Ultra) is the other viable option, and costs proportionally the same. But neither of those comes anywhere close to beating the commercial models on quality — and the per-token price of Claude, GPT or GLM, combined with their brutal inference speed, makes the math hard to justify for anyone who isn’t an enthusiast or a researcher. Expensive local AI hardware is a weekend hobby, a tool for people who need to run offline for compliance reasons, or a research playground. For day-after-day production work, right now, cloud is still the rational choice. A 128 GB Ryzen AI Max may look tempting on the spec sheet, but if the goal is serious coding agent work, it’s money badly spent.

The Qwen family: Coder vs General, distillation, and why nothing is a silver bullet

With so many different Qwens running in this rerun, it’s worth doing a more focused analysis. What I learned might surprise people who follow model benchmarks on Twitter.

Before getting to the results: what quantization is and what distillation is

These two concepts come up constantly in this discussion and they deserve a quick explanation.

Quantization is the technique of compressing the model’s weights so they take up less memory. A model trained in FP16 (16 bits per weight) can be quantized to Q8 (8 bits), Q4 (4 bits), Q3_K_M (3 bits, but with medium-sized groupings), and so on. Each step halves the size of the model on disk and in VRAM, at the cost of some loss of precision. Q8 is practically lossless. Q4 already loses something measurable. Q3 loses more. Q2 is the line where the model starts saying real nonsense. The rule of thumb is that for coding and multi-step reasoning, you want to stay at Q4 or higher. Q3_K_M is the minimum that still works for many models, and it’s what fits a 27B on the 5090 with 128k context.

The surprise from my test, and look, this goes against the consensus, is that quantization wasn’t the bottleneck here. I ran the Qwen 3.5 27B Claude-distilled in two versions: Q8 on the AMD Strix (~27 GB of weights) and Q3_K_M on the 5090 (~12 GB of weights). Both hallucinated exactly the same fake RubyLLM APIs. Q3_K_M even produced a cleaner Gemfile. The model’s limitation was in what those weights know, not in the precision they were compressed to.

Distillation is the technique of training a smaller model (the “student”) to imitate the output or behavior of a larger model (the “teacher”). The classic version is logit distillation — the student learns to approximate the teacher’s probability distributions. The modern, more popular version for coding agents is distillation of reasoning traces: you take chain-of-thought from the big model on real problems and train the smaller one to reproduce the same reasoning style.

The hype of the moment is distilling Claude and GPT into open source models. The promise is that you can have “Claude-at-home” running locally. I wanted to test this, and that’s why I added Jackrong’s Qwen 3.5 27B distilled from Claude 4.6 Opus to the benchmark. If any open source model was going to use RubyLLM correctly, this was the bet — after all, in the entire benchmark, Claude and GLM 5 are the only ones that get the API right.

What the Claude-distilled learned (and what it didn’t)

I ran the same distillation twice: once at Q8 on the AMD Strix (which blew through the 90-minute timeout), and once at Q3_K_M on the 5090 (completed in 12 minutes). Both produced the same elegant frustration.

The code that comes out looks like Claude. It has # frozen_string_literal: true at the top of every file. It has a separate Response class as a value object with explicit attribute readers. It has a clear separation between service, controller and model. It has doc comments at the top of every file. It correctly comments out things like active_record, active_job and action_mailer in application.rb. It has defensive case statements trying multiple return formats. Stylistically, it’s Claude.

Functionally, it’s a complete RubyLLM hallucination. Look at the service generated by the 5090 run:

RubyLLM::Chat.new.with_model(@model) do |chat|
  conversation_history.each do |msg|
    chat.add_message(role: :user, content: msg[:content])
  end
  response = chat.ask(message)
  Response.new(content: response.text, usage: build_usage(response))
end

Every primitive in this code is invented:

• RubyLLM::Chat.new — the constructor isn’t public, the correct entry is RubyLLM.chat(model:)
• .with_model(@model) do |chat| ... end — there’s no block API like that
• chat.add_message(role:, content:) — doesn’t exist
• response.text — the real API exposes response.content
• response.usage.prompt_tokens — the object doesn’t have that shape

This will blow up with a NoMethodError on the first request. The initializer also tries config.openrouter_api_base= which doesn’t exist on RubyLLM.configure, so the app probably won’t even boot.

The Q8 version on the AMD Strix does the exact same thing, with one difference: the entry call is RubyLLM.chat(model:, provider: :openrouter) — the entry point is right, but provider: is invented and it’s immediately followed by the same fake chat.add_message(role:, content:). Worse, the Gemfile from the 90-minute run lists gem "ruby-openai" (wrong gem!), gem "minitest", "~> 6.0" (minitest 6.0 doesn’t exist) and gem "tailwindcss" (wrong gem name, it’s tailwindcss-rails). The Gemfile doesn’t include the gem the service code itself is trying to use.

For comparison, look at the actual Claude Opus 4.6 baseline, in the same benchmark, getting it all right:

@chat = RubyLLM.chat(model: model_id)
response = @chat.ask(message)
response.content

Twelve lines in the entire service. Zero hallucination. Includes streaming via block. The distilled model produced three times the code volume and got the API wrong.

The honest reading is that distillation transferred one layer and stopped. The layer that came along was the style: code organization, comments, class structure, the order of things. The layer that got left behind was factual memory about specific libraries. That makes sense when you think about it: Claude’s reasoning traces, even when written carefully, rarely contain repeated references to chat.ask(msg).content in some obscure Ruby gem. The student only learns what the teacher repeats, and Claude never had any reason to keep whispering “use ask, not complete” throughout its chains of thought. Library API knowledge is binary recall memory, the kind that’s either in the weights or it isn’t. Decomposing that into reasoning steps is impossible because it isn’t reasoning, it’s just raw memorization.

To wrap up the practical recommendation: if you need the model to actually use RubyLLM, or any less-popular library for that matter, Claude distillation won’t save you. Use real Claude or GLM 5. The “Claude-stand-ins” in open source will fail the same way the Qwen base would, just with prettier handwriting.

Coder vs General: the surprise of the “for coding” models

Almost everyone’s instinct is that models with “Coder” in the name are the best for programming. Makes sense, they were specifically fine-tuned on code. But in the benchmark, it was exactly the opposite.

Of the three dedicated Coders, two failed catastrophically (full timeout and hardcoded mock string) and one didn’t even run properly because of an infra bug. Meanwhile, the Qwen 3.5 35B-A3B, which is the general model in the line (not the Coder), came closest to something usable: 5 minutes of execution, recognizable Rails project, and the problem is fixable in 1-2 prompts.

Qwen 3 Coder 30B is particularly disappointing. It went so far past trying to use the API that it didn’t really try at all: the controller it generated has this:

class Api::V1::MessagesController < ApplicationController
  def create
    render json: {
      response: "This is a mock response. In a real implementation, this would connect to RubyLLM with Claude Sonnet via OpenRouter."
    }
  end
end

The Gemfile lists gem "ruby_llm" but nothing imports it. The service layer is nonexistent. The model decided it was easier to return a fake string and call it a day. That’s Tier 3 garbage in a way no correction prompt fixes — you have to tell it to start over.

Qwen 2.5 Coder 32B is even worse: 90 minutes running, zero files. The 1.8 MB opencode-output.ndjson shows the model spinning without managing to write anything. It probably got stuck in a planning loop without ever calling the write tools. Total slot waste.

Why did the “Coder” Qwens do so badly? My read is that the coding-specific fine-tuning they got was trained on more isolated problems (Codeforces, Leetcode, short snippets), far from agentic flows with long-running tool calling. The general Qwen 3.5 35B-A3B has broader training and handles the orchestration part better. The popular intuition “Coder = best for coding agent” is wrong for this kind of task. The use case where Coders shine is “complete an isolated function,” which is exactly what they were trained for, and that’s a tiny fraction of what a coding agent does day to day.

The question I wanted to answer

It was this: running locally on the 5090, which Qwen model is worth the 1-2 correction prompts to deliver code that works?

The honest answer is: only Qwen 3.5 35B-A3B, and maybe the Claude-distilled if you don’t mind spending 12 minutes more.

• Qwen 3.5 35B-A3B on the 5090: 5 minutes, correct entry point (RubyLLM.chat(model:)), errors on the subsequent calls. Realistic total until it works: in the 15-20 minute range with 1-2 follow-ups. Beats cloud OSS on cost.
• Qwen 3.5 27B Claude-distilled on the 5090: 12 minutes, deeper hallucination (entry point is invented too). Realistic total: 25-30 minutes with 2-3 follow-ups. Still competes on cost, and loses on absolute time to the real Claude.
• The others (Coder 30B, Coder 2.5 32B, 3 32B): don’t pay back the correction time. Each one has a structural problem that calls for a full rewrite from scratch.

For folks with hardware in this category who want to escape Anthropic vendor lock-in, it now works. It didn’t work on the 5090 from last year, and forget about it on the Strix Halo. In 2026, on NVIDIA Blackwell, with the right model, it works. For folks with low-bandwidth hardware (LPDDR5x, DDR4, DDR5), it’s still a waste of time: the clock alone takes down any plan to make this practical.

Qwen 3.6: what changed from 3.5

We tested two flavors of Qwen 3.6: the 3.6 Plus (cloud, OpenRouter, free) and the 3.6 35B (local, NVIDIA 5090, Q3_K_M).

Qwen 3.6 Plus (cloud) completed in 17 minutes with 88,940 tokens and 9/9 on the artifact checklist. Completed fast and for free. But the generated service uses chat.add_message(), a method that doesn’t exist in RubyLLM. First message works, second one crashes. Same problem as 3.5.

Qwen 3.6 35B (local, 5090) is more interesting. Completed in 4.7 minutes at 240 tok/s, 169 files, correct entry point RubyLLM.chat(model:, provider:) and correct chat.ask(message). The bug is subtler: it returns response instead of response.content and doesn’t do history replay. One-line fix. That’s a real improvement over Qwen 3.5 35B-A3B (which hallucinated add_message and complete). It’s the cleanest Qwen result we’ve seen.

The gap shrank but didn’t close. The 3.6 35B local is closer to working than any previous Qwen — the bug is a forgotten .content, not an entirely invented API. But it still doesn’t do multi-turn out of the box. In practice, Qwen 3.6 35B local moves up from Tier 3 to Tier 2: it’s the local open source model that comes closest to delivering correct code on the first try, one fix away from working.

The Deep Code Review: Sonnet vs GLM 5 vs Gemini vs Kimi vs MiniMax

The tables above measure structural completeness. But does the project work? I did detailed code review of the models that completed the benchmark.

Claude Sonnet 4.6 — works and is the most complete. Synchronous responses via Turbo Stream. Chat history persisted in a session cookie with full replay of previous messages on every request. Correct LLM mocking in the tests with mocha (30 tests in 328 lines). LLM logic extracted into a separate LlmChatService. Views decomposed into 9 partials. Minor problems: duplicated model constant, leak in the auto-resize event listener. None are blockers. Of the generated projects, it’s the closest to something you’d actually put into production.

GLM 5 — works, but it’s the bare minimum. Uses the correct API (RubyLLM.chat(model:) then .ask()), does mocking with mocha in the tests. But the project is way leaner than Sonnet’s: 21-line controller (vs Sonnet’s 52), no service layer (LLM logic inline in the controller), no chat history persistence, every message handled in isolation. The first message works, but the app doesn’t keep conversation context, so you can’t have a multi-turn dialog. The tests exist (7 methods) but they’re skeletal: ruby_llm_test.rb only checks that the module is loaded, chat_flow_test.rb is a copy of the controller test. The Dockerfile, on the other hand, is the best of the four: multi-stage, non-root, jemalloc. But as a chat app? It’s more of a proof of concept than something functional. Funny detail: the README says “Powered by Claude Sonnet 4” instead of the model that actually generated the project.

Gemini 3.1 Pro — the fastest, but trips on the API. Completed in 14 minutes, the fastest along with MiniMax. The Rails code itself is well written: uses Rails.cache with session ID and a 2-hour expiration to keep state (instead of a database), Turbo Streams nicely integrated, Stimulus controller for auto-scroll, and the Dockerfile is the best of the group (multi-stage, non-root, jemalloc). The problem is the usual one: it uses RubyLLM::Chat.new() instead of RubyLLM.chat(), and calls add_message() which doesn’t exist. The app boots, Docker runs, the health check passes, but the first chat message returns 500. The tests (5 methods) mock with a FakeChat that replicates the wrong signature, so they pass. It’s frustrating because the rest of the code is the most “Rails way” of the non-Anthropic models. Fixing it would be 3 lines, but the benchmark measures what comes out the first time.

Kimi K2.5 — ambitious but broken. Tried the most sophisticated architecture: ActionCable streaming, configurable models, dual Dockerfiles, 37 tests in 374 lines. Problem: the streaming depends on ActionCable, which is commented out in config/application.rb. The return unless defined?(ActionCable) guard makes the method do nothing. The assistant never responds. The Stimulus controller has a scope bug: submitTarget references a button outside the controller’s subtree. Thread-unsafe storage with a hash in a class variable. Kimi wrote more tests than any other model (37), but none of them mock the LLM calls — so the tests pass without proving any of the functionality works.

Grok 4.20 — fast and wrong. It was the fastest in the entire benchmark: 8 minutes, 412 tok/s. Except it was fast because it cut corners. The prompt explicitly asked for the ruby_llm gem, and Grok ignored it. It went straight for OpenAI::Client from the ruby-openai gem pointing at the OpenRouter URL. Technically the first message comes back, so yeah, it “works.” But it’s the same trick as Step 3.5 Flash and Qwen 3.5 122B: skip the part that was actually being tested. No history, 33-line controller calling the HTTP client by hand, two tests, no real mocks. It was fast because it did less than what was asked.

MiniMax M2.7 — looks right, crashes. Calls RubyLLM.chat(model: '...', messages: [...]) — that signature doesn’t exist. No message persistence. Duplicated HTML (DOCTYPE inside the layout). Committed master.key. And the tests? They mock the wrong API, so they pass but they don’t prove anything.

Code review summary: