See where the world’s AI models actually come from.
3,484 tracked models, traced to the country of the organisation that built them. Two countries account for most of the volume — but volume is not the same as the frontier, and the tail is where the interesting labs are.
- United States1,748 models tracked
Claude Fable 583.0LB
- China829 models tracked
Kimi K379.2LB
- United Kingdom178 models tracked
Stability AI · 8 recent
- Canada106 models tracked
Command R+1226Elo
- South Korea83 models tracked
LG AI Research · 15 recent
- France76 models tracked
Mistral Medium 3.51427Elo
- Germany74 models tracked
Black Forest Labs · 6 recent
- Japan53 models tracked
Tokyo Institute of Technology · 2 recent
Each country shows its highest-scoring model, not its most prolific lab. Counting models rewards whoever publishes most research: in the United States that is a 64 / 60 / 57 photo finish between Google DeepMind, NVIDIA and OpenAI, while Anthropic ships 17 models and holds the top score. Volume and quality are different questions and this list answers the second one.
See which model to start withNot drawn: Multinational (10) — listed rather than dropped, because a value we cannot place is still a value we hold.
Release velocity
Models entering the registry per quarter. The shape is the point — this is a rate, not a set of categories.
Complete quarters only; the quarter in progress is excluded rather than drawn as a cliff. A soft tail in the most recent quarters is expected regardless — Epoch catalogues models after release, so the newest periods keep filling in for a while.
Discover
Where to start, by task, industry and language.
Models
The full registry, filterable by what a model accepts.
Rankings
Capability, provenance and recency across the field.