Hugging Face Model Hub
by Hugging Face, Inc.The Model Hub is infrastructure the open AI ecosystem would struggle to function without. It hosts hundreds of thousands of models — language, vision, audio, multimodal — from Meta, Mistral, Google, independent labs and individual researchers, with standardised hosting, versioning and documentation. When a new open-weights model is released, this is where it appears.
Its real contribution is standardisation. Before it, using someone else's model meant finding their repository, deciphering their loading code and reconstructing their environment. The transformers library plus the Hub reduced that to a few lines that work identically across thousands of models. That consistency is why the open ecosystem moves as fast as it does.
Model cards are the other piece: standardised documentation covering intended use, training data, evaluation results, limitations and known biases. Compliance varies — plenty of cards are thin — but establishing the norm that a model ships with documentation was a meaningful contribution to the field's practices.
Licensing is the thing users most often get wrong. 'Open weights' is not one thing. Some models are Apache or MIT; others carry custom licences with commercial restrictions, usage limits or acceptable-use policies. Llama's licence is not Mistral's. Anyone deploying commercially needs to read the specific licence rather than assuming availability implies permission.
Public hosting is free. Paid tiers cover private repositories, hosted inference and enterprise features. It suits ML engineers, researchers and anyone building on open models — which is most of the field outside the closed frontier labs.