Anyone can download an open-weight model, but that is not the same as open source. The license decides what you may do with the files. Each model page on this site shows the license name; this guide explains the common ones. It is general information, not legal advice, and licenses can change between versions of a model, so read the license on the model’s own page before building on it.
The common kinds
| License | In short |
|---|---|
| Apache 2.0 | Use, modify and sell products built on it, including commercially. Keep the notices. Includes a patent grant. Many Qwen, Mistral and Gemma releases use it. |
| MIT | Even shorter and just as permissive: keep the copyright notice. No explicit patent grant. |
| Llama community licenses | Free to use, including commercially, with conditions: a very large user-count threshold above which you need Meta’s permission, an attribution requirement, and an acceptable-use policy. Terms differ by Llama version. |
| Other custom licenses | Some labs publish their own terms, often with an acceptable-use policy, limits on using outputs to train competing models, or non-commercial restrictions. Read them. |
| Research or non-commercial | Fine for personal experiments. Not for a product or paid work. |
Questions to ask
- Is this for personal use or for a product? Personal use is rarely restricted. Products need a permissive or an explicitly commercial license.
- Do I need to keep a notice or say “Built with…”? Apache 2.0, MIT and some custom licenses require this.
- Will I use outputs to train another model? Some licenses forbid it.
- Is it gated? Some models require you to accept terms on Hugging Face and sign in before downloading. That is a download step, not a different license.
- Did I fine-tune it? A fine-tune usually inherits the base model’s license.
Open weights also say little about training data, which most labs do not publish. If that matters to you, look for models that document their data. Back to choosing a model.