About this site

A bilingual directory and benchmark comparison site for AI. Entries are split into China and Global modules and cover language/multimodal, image, video and speech/audio models plus AI agent tools. Each entry lists official links, a bilingual overview, version lineage, openness and multi-dimensional scores. Scores are labelled as vendor-reported, independently measured or human preference; missing values are only ever “not tested”, “not disclosed” or “not applicable” — never zero.

Every number on the site comes from an official page or institutional leaderboard that opens and can be quoted verbatim; the citation sits next to the number and under Practical info. Values we cannot source are left blank — never zero, never estimated, never borrowed from another model. Where sources disagree on a field, both readings stay in the entry with their citations and readers judge for themselves. Leaderboard data is re-checked against the official snapshot after each crawl, and any anomaly is recorded in the update log.

Per-field sourcing conclusions are in the verification report shipped with the site. · Last updated: 2026-11-08

Scores and prices come from third-party leaderboards and public vendor materials and may be stale or wrong; always confirm on the vendor's site. Logos and trademarks belong to their respective owners and are used here for identification only. The naturalness note is subjective and qualitative, not a numeric rating. Sources for every figure can be inspected on detail pages.

Email us if you spot an error. Messages are handled during the weekly maintenance window; we do not promise individual replies. Verified fixes ship with the next publication.

Scores and leaderboard data come from third-party sources. Each source's licence and attribution terms are registered when it is enabled; where attribution is required it appears both next to the data and here. Sources whose terms are unconfirmed are link-only — never reproduced as tables — and are labelled as such here.

Per-source licence and attribution registration is recorded in the verification report shipped with the site. Unregistered sources are never reproduced as full tables.

All cited sources (by citation count)

Every source actually cited in this index, deduplicated by domain and publisher name, ordered by citation count. Each score keeps its own source link in the technical layer of the entry page.

All cited sources (by citation count)

Every source actually cited in this index, deduplicated by domain and publisher name, ordered by citation count. Each score keeps its own source link in the technical layer of the entry page.

Data refreshes twice daily at about 08:00 and 20:00 Beijing time; scheduled jobs may run 5–30 minutes late. If a fetch fails, the previous values and their original date stay visible — never a blank page or an error screen. New models go live after manual review.

Backup trigger: an external scheduler pings GitHub at 08:05 and 20:05 daily, so the platform schedule is not the only trigger.

Sources at a glance

Each source is listed with its coverage, cadence and latest result. Addresses and licences are registered in data/crawl/sources.json — adding a source only touches that file.

Last updated: 2026-11-10

Currently using the public-page semi-automatic branch (no free key configured). Those three fields update less reliably; all other fields are unaffected.