Open-source
Open Weights
Models whose trained weights are publicly downloadable and runnable (Llama, Mistral, DeepSeek), but without released training data or code — less than fully open source.
Open weights refers to models whose final trained parameters — the weights and biases of the neural network — are publicly available for download. Anyone can run them on their own hardware, study how they behave, and fine-tune them further, without depending on the originating lab’s API.
Open weights are distinct from fully open source. An open-weight release ships only the learned parameters; it does not necessarily include the training data, training code, or detailed methodology, so full reproduction is impossible. The Open Source Initiative therefore treats it as a compromise — more open than closed models, but short of the “open source AI” standard. Many releases also carry restrictive licences (for example the Llama Community License) that limit some commercial uses, which is why “open weights” and “open source” are not interchangeable.
In 2025–2026 this category became a competitive core of the ecosystem: Meta’s Llama, Mistral, DeepSeek, and Qwen form the open-weight frontier, enabling local inference, private deployment, and research outside hyperscaler clouds.