News · 2026-08-24

A 22-gigabyte local coder matched Opus on a 25-problem slice

TielCoder, a 4-bit re-quantization of the Ornith-1.5-35B-A3B mixture-of-experts model, fixed 12 of 25 problems from a live software-issue benchmark in its maintainer's tests -- the same count as Claude Opus 4.6 at medium effort. The benchmarked file is a 22.4 GB download that the model card lists as fitting a 24 GB consumer graphics card. Whatever else is true, a frontier-adjacent coding result no longer requires a data center.

Key facts

• The benchmarked tier, Tiel-Coder-35B-A3B-UD-Q4_K_XL.gguf, is a 22.4 GB download; the card's fit table calls it "snug on 24 GB, comfortable on 32."
• It fixed 12 of 25 problems on a SWE-bench-Live slice, matching Opus 4.6 (medium), four more than its own base model.
• It scores 73.7 on MMLU-Pro against a sibling model's 84.0, at the same 4-bit precision.
• Published on Hugging Face under an MIT license, from base model Ornith-1.5-35B-A3B.

The interesting thing about this release is not the benchmark. It is the shape of the tradeoff, which the maintainer describes with more candor than most vendor announcements manage.

Some background on what "4-bit quantization" buys you. A model's weights are normally stored at 16 bits per number. Quantization rounds them to fewer bits -- here, about four on average -- which shrinks the file roughly fourfold and lets it fit in a graphics card you can actually buy. You lose some accuracy in the rounding. The craft is in losing it in the right places, which is what an importance matrix does: it measures which weights matter most during real use and spends precision on those. The full tier list here runs from a 12.3 GB build for 16 GB cards up to a 38.5 GB reference build, and the card is explicit about which one to take: "start here -- the benchmarked tier." Ground Truth's explainers on quantization and model file formats cover the mechanics.

The underlying model is a mixture of experts, which is why a 35-billion-parameter model runs fast on consumer hardware: only a fraction of the network activates for any given token. The card also notes a hybrid state-space and attention design with only two key-value heads, so the memory used by the growing context is far smaller than the file size would suggest -- meaning the 22.4 GB figure leaves more room to work than it looks like it should. See mixture of experts and KV cache for why that matters.

The results are stated with their scope attached. On 25 problems from SWE-bench-Live, a benchmark drawn from recent real GitHub issues rather than a frozen historical set, Tiel fixed 12 -- level with Opus 4.6 at medium effort, four ahead of its own base model, three ahead of a sibling quantization, and four ahead of Sonnet 5 at medium. It was also faster and steadier: an 8.6-minute median per attempt without the long tail of expensive failures its sibling has. On multi-turn conversation, scored across 114 conversations each, it beat both its base and its sibling, earning the gain on answer quality rather than by asking more clarifying questions.

And then the cost, stated plainly. "It is also cheerfully bad at trivia," the card says. "Pick it for work. Pick something else for exams." The number behind that is 73.7 on MMLU-Pro against a sibling's 84.0. Most of the gap is inherited from the base model rather than introduced by the quantization, and the remainder is attributed to a chat template that buys shorter answers -- an explicit trade the build was made to take.

Why it matters: this is what commoditization looks like from below. OpenAI just cut Sol's price by a third on output, and part of the pressure comes from the fact that a developer with a 24 GB card can now run something that fixes real repository issues locally, at zero marginal cost and with no data leaving the machine. Meanwhile self-hosting a genuinely frontier open model still takes a whole node, so the useful question is not "can I run the biggest model" but "what is the smallest model that clears my bar." Ground Truth previously covered how the Ornith-1.5 base model writes and grades its own training problems.

Two caveats worth holding onto. First, 25 problems is a small sample, and the card itself warns that small differences on that benchmark should be treated as noise -- this is the maintainer's own evaluation run, not an independent leaderboard. Second, a maintainer note that reads as a warning sign turns out to be the opposite: the base model originally shipped an untrained speculative-decoding head, which this build omitted as dead weight after measuring its weight statistics. Ornith fixed it with a re-upload on August 23, 2026, and the trained head now ships in a separate repository for runtimes that use it. That level of documented forensic detail about a file's contents is rare, and it is a better signal of care than any benchmark on the page.

Primary source, verified: read the paper →

Ground Truth. (2026, August 24). A 22-gigabyte local coder matched Opus on a 25-problem slice. Ground Truth. https://groundtruth.day/news/a-22-gigabyte-local-coder-matched-opus-on-a-25-problem-slice.html

@misc{groundtruth:a-22-gigabyte-local-coder-matched-opus-on-a-25-problem-slice,
  title  = {A 22-gigabyte local coder matched Opus on a 25-problem slice},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/a-22-gigabyte-local-coder-matched-opus-on-a-25-problem-slice.html}
}

Topics: open-weight-models · quantization · local-inference · coding · mixture-of-experts · gguf