We read only the safetensors headers of 12,974 Hugging Face repositories (names, shapes and storage of every tensor, never the weights) and recovered each model's architecture from them. This page shows what widely used models are made of, how that has changed by year, and where their metadata is wrong.
Type any model from the census. See what it is made of, where it sits among popular models, and which devices it fits on.
Leading providers' releases, the most downloaded language models, Thai and multilingual models, and models submitted by users. Each is read from its headers, assigned the smallest device that holds it at 4-bit with an 8k context, and given five metadata checks. Submitted models receive a badge for their model card.
Popular repos: the most downloaded, the most downloaded original releases, and recent releases from leading providers. Test fixtures (tiny random-weight repos) are left out.
What thousands of checkpoints show about how models are built, copied, stored and run. Shares come with a 95% confidence interval, and changes over time with a trend test. The method is in the paper; the code is under research/hub-census.
The most downloaded readable models in the census. Click a name to open its full anatomy.
Params: logical parameters (quantized storage unpacked). Hub: ✓ when the Hub's own count agrees within 0.1%, otherwise our count ÷ the Hub's. KV/token: bytes of KV cache per token at 16-bit.
For every repo, the dataset gives status, popularity, the architecture recovered from the weights, the values its config.json declares, the Hub's own parameter count and the raw manifest, so the results can be recomputed. The code and the fixed recovery rules are public.