Methodology
Five kinds of entry
Everything on Explore is one of five kinds. Each is judged its own way, and every result links back to what produced it.
- TopicsA curated page on one field or theme that gathers its benchmarks, tasks and showcases.Not scored. It shows the results of what it gathers, each with its own source.
- BenchmarksA fixed set of questions written for a professional field, versioned so every run answers the same questions.Multiple-choice answers are checked by a program. Short answers are graded by a judge model against the reference answer and rubric, and the judge is named on every run.
- TasksA real job with its materials, the deliverable it asks for, and acceptance checks.The work passes or fails each acceptance check. The result is how many checks it passed.
- ShowcasesA brief for a work worth showing: a 3D model, an image, a video or an interactive page.People look at the works side by side and vote blind, plus any checks the brief sets out.
- ComparisonsTwo to four models on one benchmark, same questions and same judge, published as one record.Each model is scored as in a benchmark run, and the page shows where their answers differ.
Community tiers
A tier describes how much of your work the community can check. It is not a score, and it is not for sale.
- Member — at least one verified work: a run the platform regraded, a result someone independently reproduced, or an accepted delivery.
- Regular — ten or more verified works, on at least two different benchmarks or tasks, over at least thirty days.
Only work that can be checked counts. Followers don’t count. Self-reported links don’t count: they stay on your profile with their own label, but they never move your tier. The verified works count on a profile uses the same definition, so it includes reproductions and accepted deliveries, not only regraded runs.
Tiers recompute automatically, and every change is logged. An account goes under review when two of its runs fail regrading within thirty days. While it is under review, its tier is hidden; the tier comes back when the review resolves.
Where the other numbers come from
BenchClub runs are the only numbers we measure. Everything else is collected from these sources, shown with its source and date, and never mixed into run results.
| Source | What we show |
|---|---|
| Hugging Face | Which open models exist, their size and license, downloads, likes, trending score, and what people built on them |
| LMArena | Arena scores by category, measured on the vendor's served model, not a local build |
| OpenRouter | API tokens used each day for its 50 most used public models |
| Hacker News | Posts about a model with 20 or more points in the past year |
| YouTube | The most viewed videos of the past 90 days about the 40 most downloaded model families |