Using the Index

Rank by

Overall ranks models scored by boards from two independent publishers. A category (Agents, Coding, Reasoning, Knowledge, General, Human preference) or a domain inside it ranks every model a recognised board has measured there; rows placed only by report figures carry a ✱ after their number.

When no recognised board measures a domain yet, a note above the table says so: every figure there is from one publication, a comparison set its publisher chose. Read those positions as “within that set”.

Filters

Filter	What it does
Search	Model name or organisation
Model line	Claude, GPT, Qwen, K2, … read off the name
Size	Small ≤ 15B · Medium 15–100B · Large 100B–1T · Very large > 1T · Undisclosed. Counted from the weights on Hugging Face for open models, read from the name otherwise; closed models are undisclosed, never estimated
Coverage	Minimum number of recognised boards; “Any coverage” includes report-only models
Include reports ✱	Off: report figures leave every score and report-only models are hidden
Must include	Only models scored by the chosen sources

Reading a row

• Index — the Overall score. ~55✱ is an estimate; ≤ / ≥ mean the model sat outside the range of the models it could be compared with.

• Sources — one mark per recognised board (filled = scored; hover for the figure), a ✱ with a count for reports, and n/N coverage.

• provisional — fewer than two independent publishers.

The model card

Open a row for: facts about the model (organisation, size, open weights, context, first listed), scores by category → domain, how to call it (OpenRouter id, vendor site, weights) and every source figure with the exact label the source printed.

Sharing

Every filter lives in the URL. Share on X / LinkedIn / Copy link share exactly the current view; the link previews as a card of that view’s top ten. Export chart downloads the same top ten as a PNG.