PostWeights Not Walls
How to read the scores on this site
Every score here is labelled claimed, reported or re-run. What each label means, whose numbers they are, and what we do not do yet.
Every score on this site carries one of three labels: claimed, reported or re-run. The label tells you who produced the number, and so how far to trust it. We never average scores of different kinds or blend them into one ranking. The full definitions are on the method page; this post shows what they look like with real rows from our data, as of the snapshot of 19 November 2026.
Claimed
A claimed score comes from the model's maker, in the model card or a paper. It may well be right. But it was measured by the people with the most to gain from a good number, with settings they chose, and usually without logs.
Example: the card for Mellum2.1-12B-A2.5B-Thinking-MTP, which JetBrains published on the Hub on 30 October 2026, lists 17 scores, among them 64.6 on GPQA Diamond and 47 on SWE-bench Verified. We have no independent result for this model, so the maker's claim is all there is, and the page says so.
Few makers publish scores in a form a program can read. In the snapshot, 7 of our 390 reviewed models do. For the others, any claims sit in the prose of the card, and we do not copy them by hand.
Reported
A reported score was published by someone other than the maker, usually a public leaderboard, and collected by Epoch AI. We show the leaderboard's name and link to it. Reported scores are independent of the maker, but we cannot check how each leaderboard ran its tests.
Example: on Terminal Bench, the Terminal-Bench 2.0 leaderboard lists gpt-oss-120b at 18.7% with the Terminus 2 agent, and at 14.2% with Mini-SWE-Agent. Same model, same test, different harness around it: 4.5 points apart. A score without its setting tells you little, so we show the setting next to every score.
Settings inside the model matter too. On the WeirdML leaderboard, the same gpt-oss-120b scores 48.2% at its high reasoning setting and 41.9% at medium. Those are two results, and we list them as two.
Re-run
A re-run score comes from someone independent who ran the test themselves and published the logs. Today every re-run score on this site is a run by Epoch AI, and the badge says "Re-run by Epoch AI".
Example: Epoch AI ran gpt-oss-120b, at its high reasoning setting, on GPQA diamond on 11 December 2025. It scored 75.8%, plus or minus 2.7. The log of the run is public: you can read each question and each answer. The plus or minus is the standard error. Two models a couple of points apart on this test may not differ at all.
Whose numbers these are
None of them are ours. Claimed scores are the makers', reported scores are the leaderboards', and re-run scores are Epoch AI's. Our part is to collect them, label them and attach each one to the right model.
That last step is where a site like this can do most harm. Epoch AI names models in its own way, and we match those names to Hugging Face repos. Where a match is uncertain we leave the result unattached rather than guess, because a score on the wrong model is worse than no score. In the snapshot, 135 of our 390 reviewed models have at least one reported or re-run score. The rest have none yet, and their pages say so.
What we do not do yet
- We do not run any evaluations ourselves. Our own re-runs, on open test harnesses with public logs, are planned and not built. When they start they will be labelled as ours, separately from Epoch AI's.
- We do not test providers. The providers table shows the precision each provider states and the speed the Hugging Face router measures. Whether a provider serves the real weights at that precision is not tested yet.
- We do not keep score history, so we cannot show how a model's results changed over time.
- We do not combine scores into one overall number.
If a number is wrong
Tell us. The content policy says how, and corrections are made in the open. Scores from Epoch AI are used under CC-BY 4.0; each row links to its own source.