The model register

What is in here. Models published by 25 tracked organisations, plus any model passing 50k downloads. Parameter counts are exact, from the safetensors index, never rounded. Context length is max_position_embeddings from the model’s own config.json — the architectural maximum, which is not always what a provider serves or a vendor advertises. A further 93 records are quantisations and conversions of other people’s releases (GGUF, MLX, AWQ and similar). They are kept in the data and excluded from the table, because they are repackagings rather than releases — they are listed below, and in the API and the CSV.

Reading the capability columns. These are derived from each model’s own chat template — the Jinja that defines how a conversation is serialised for it. A template that handles tool_calls is evidence the model was trained to emit them, which is a stronger signal than a vendor claim. ● evidence found. ○ template read, none found. ◇ no chat template published, so we cannot say — which is not the same as unsupported.

The Neural Net. (2026). The Neural Net model register [Data set]. https://theneuralnet.com/models

@misc{theneuralnet_model_register,
  title  = {The Neural Net model register},
  author = {{The Neural Net}},
  year   = {2026},
  url    = {https://theneuralnet.com/models},
  note   = {CC BY 4.0}
}

<a href="https://theneuralnet.com/models">The Neural Net model register</a>