The most useful question about a local model is not which name has the biggest halo. It is narrower: what can my own machine actually do?
On blindtest.ai, the best local simulation result so far is qwen3.5:9b at 36% runnable, 31 of 85 simulations passing a runtime check. The frontier reference models sit in a different band: Claude Sonnet 5 is 100% (85/85), Claude Fable 5 is 100% (46/46), Claude Opus 5 is 100% (60/60), GPT-5.5 is 95% (81/85), Claude Haiku 4.5 is 81% (68/84).
That gap is the opening fact because it is concrete and checkable. A simulation either renders, animates, and avoids a JavaScript error in a real browser, or it does not. The local field is not empty: qwen3.5:9b, mistral-small:24b (33%, 28/85), and qwen3:14b (28%) all produce runnable simulations in a visible share of tasks. But the distance to the frontier references is large, and it shows up under runtime verification, not preference voting.
Text tells a different story. On blind rubric-scored text rounds, gemma4:26b leads the local field at a 9.61 average across 31 rounds, with qwen3-coder:30b (9.34) and qwen3.5:9b (9.21) close behind. Local models can be competitive at writing. Where that competence transfers, and where it stops, is the rest of this report.
Method
blindtest.ai runs one-shot tasks: one attempt per model, no retries, no repair pass after a bad answer. Answers are published unedited, including typos, wrong answers, refusals, timeouts, and errors.
The site covers 159 rounds total: 130 blind rounds (r31-r160) plus 29 archive rounds (r1-r30), across 8 disciplines. The field is 11 local open-weight models running on the author's own consumer-GPU cluster: gemma4:26b, mistral-small:24b, llama3.1:8b, qwen3-coder:30b, qwen3:14b, qwen3.5:9b, qwen3:8b, deepseek-r1:14b, deepseek-r1:32b, qwq:32b, and command-r:35b. Frontier references are tested separately and not scored on the same axis: Claude Sonnet 5, Claude Opus 5, Claude Fable 5, Claude Haiku 4.5, and GPT-5.5 (via Codex).
Code and simulation rounds are not judged by reading the code. The output is loaded into headless Chrome and checked at runtime: does it render, does it animate, does it throw a JavaScript error. A plausible-looking file can still be static, blank, broken, or too large for the requested format, and only a runtime check catches that.
Text rounds are scored blind against a written gold answer and rubric, with model identity revealed only after judging. That removes the name effect: an answer is not credited for a famous logo or penalized for a small local one.
Harness failures are excluded from model scores; if infrastructure breaks, the model is not blamed for it. There is no voting and no composite score blending unlike tasks into one headline number. Text rubric score and simulation runnable-rate stay separate axes, and every number links back to the raw answer that produced it.
One line of context: in a period where aggregated crowd preference has been questioned in the LMArena "Leaderboard Illusion" debate, blindtest.ai is best read as a verifiable primary source rather than a rebuttal of any other benchmark.
Findings by discipline
Simulations / code
qwen3.5:9b leads the local field at 36% runnable (31/85), followed by mistral-small:24b at 33% (28/85) and qwen3:14b at 28%. The frontier references are markedly higher on the same runtime check: Sonnet 5, Opus 5, and Fable 5 all at 100% (85/85, 60/60, 46/46), GPT-5.5 at 95% (81/85), Haiku 4.5 at 81% (68/84). GPT-5.5's four misses include a binary min-heap animation that throws a JavaScript error at runtime - even at the reference tier, "looks right" and "runs right" are different claims.
The failure taxonomy is mechanical, not impressionistic: OK (passes), ERR (JS error), STAT (renders but nothing moves), BLANK (empty canvas), HANG (freezes the tab), LIM (character budget exceeded), FORMAT (markdown fence instead of file contents). It exists because a generated file can look close in a text editor - a canvas element, functions with animation-shaped names - and still fail the moment a browser actually runs it.
qwq:32b is the sharpest illustration of that gap between reading and running. It is only 8% runnable on simulations (7/85), and 42 of its attempts earned the STAT verdict: it renders, then nothing moves. That is a distinct failure mode from a syntax crash or a blank page - simulation-shaped output without simulation behavior.
Text quality
As of 45 scored rounds, gemma4:26b leads the local text field at 9.61 (31 rounds), followed by qwen3-coder:30b at 9.34, qwen3.5:9b at 9.21 and qwq:32b at 9.10. The bottom of the field is llama3.1:8b at 6.58 and command-r:35b at 7.25. These are blind scores - the model name is hidden until after judging - so the result reflects whether the answer satisfied the gold answer and rubric, not brand recognition. The ranking moves as new disciplines arrive: the latest creative-marketing rounds enforce strict output contracts (exact word counts, sentence counts, character limits, banned-word lists), and qwq:32b and deepseek-r1:32b both lost their earlier near-tie with gemma4:26b to precisely those contract violations.
gemma4:26b is the throughline of this report. It is the text leader, yet it almost never delivers a runnable simulation through the harness: its reasoning consumes the token budget before an answer forms, honestly logged as harness-excluded rather than scored as a failure (0 of 1 measurable). qwq:32b shows the same split from the code side - a top-tier text average of 9.10, and only 8% runnable on sims with 42 STAT verdicts. A local model's writing score does not predict its simulation behavior, which is exactly why the two axes are kept separate rather than blended into one number.
Structure & data
In rounds r146-r150, JSON repair was solved error-free by 10 of 10 models - a clean result for that task shape. CSV and extraction tasks were also nearly perfect on content, but exposed a different problem: 5 and 6 models respectively violated an "output ONLY" instruction by wrapping the answer in markdown fences. That is a formatting failure against a strict output contract, not a comprehension failure, and it would still break a downstream parser even though a human reader would barely notice it.
Table arithmetic, which included a sum/max task with a cancellation (void-row) trap, failed for 3 of 11 models - command-r:35b, llama3.1:8b, and mistral-small:24b either miscalculated or fell into the trap. Three separate findings, three separate failure modes: repair solved, output discipline uneven, arithmetic-under-a-trap still catching a quarter of the field.
Language
In rounds r141-r145, register-switching and plain-English rewriting were handled by almost all models - the broad edge of language competence is not where this field struggles. Idioms were more revealing: the phrase "We dropped the ball" was translated literally by two different models, deepseek-r1:14b producing "wir haben den Ball verloren" (we lost the ball) and qwen3:8b producing "die Kugel verloren" (lost the bullet/sphere). Both are grammatical German sentences that miss the idiom entirely.
A false-friends sentence, built around eventuell/aktuell/sensibel/Gift, was handled correctly by 9 of 11 models - a different test than register or plain-English rewriting, since it asks a model not to over-trust surface resemblance between languages.
What this means for "what can my own machine do"
The answer splits by discipline rather than by model tier. For text, the local field can be competitive under blind rubric judging: four models score above 9 (gemma4:26b, qwen3-coder:30b, qwen3.5:9b, qwq:32b). For simulations that must survive a runtime check in a browser, the best local runnable-rate is 36%, against a frontier range of 81-100% on the same kind of task - not a formatting difference, but the difference between an answer that can be judged by reading and one that has to work.
Competence also does not move as a block. gemma4:26b leads text and is near-zero on measurable simulations. qwq:32b scores above 9 on text and fails 42 times with a STAT verdict on sims. qwen3.5:9b leads local simulations while sitting third on text. The practical conclusion for someone choosing a model to run locally is not "which model wins" but "which model can do this kind of work, on this machine, today."
Limits of this measurement
These are blindtest.ai's own tasks, not an independent benchmark suite, and they should not be described as one. Each task runs once: there are no retries and no variance estimate, which keeps the record close to a real one-shot user experience but means the site does not measure run-to-run stability. Some categories are still small: the text results above are stated as of 45 scored rounds, and the simulation denominators differ by model (46 for Fable 5, 60 for Opus 5, 84 for Haiku 4.5), which is why those denominators stay visible rather than being reduced to a single percentage.
The methodology may change over time. That is not a weakness as long as changes stay explicit and the raw evidence stays public. The constraint that matters is already in place: infrastructure failures are separated from model failures, text and simulation axes are not blended, and every number stays traceable to the answer behind it.
Read the evidence
The full ranking is at /leaderboard/, which keeps rubric score and runnable-rate as separate axes and shows the runnable-rate trend per 10-round band for each model. For direct comparisons, /compare/ gives a 1:1 view of any two models. For the underlying evidence, the /round/rNN/ pages show the raw answers, and simulations run directly in the browser - the numbers above are not a blended score, they are a path back to the actual answer a model gave on the actual task.