Open Weights
With open-weights models, the trained model weights are published, so you can run and adapt the model yourself. That is not automatically the same as open source.
How it works
A model consists of its weights, the numerical values learned during training. With open-weights models they are publicly available: you can download them, run them yourself and adapt them. Well-known examples are models from the Llama, Mistral and Gemma families. Open source in the strict sense would be more — namely open training data and open training code — and that is not the case for many models.
A practical example
A company runs an open model on its own server and uses it to process sensitive documents without them going to a cloud provider.
What you should know
- Licenses differ: check whether commercial use and adaptation are allowed.
- Running it yourself also means being responsible yourself: hardware, operation, updates and security are on you.
- The most powerful models are often available only as a cloud service. Open models are still enough for many tasks.
- Image and video models can have open weights too, for example Z-Image or FLUX 3 Image Klein. Our overview of AI image and video models shows which ones.
- Which open language models exist, how they perform and what hardware they need is shown in our comparison of open AI models.
Note: this information is not legal advice.
Matching tools