What does it mean that an AI model is open source?

An AI model is open source when it grants the four freedoms to use, study, modify and share it, together with what you need to make changes: detailed information about the training data, the complete code for training and running the model, and the model weights. This is the standard set by the Open Source Initiative in its Open Source AI Definition 1.0, published in October 2024, which also requires all of this to be available under OSI-approved terms.

• Open source vs open-weights: many models described as "open" only let you download and run the weights. The training data and full training code are not published, or the licence adds restrictions.
• Restrictive licences: Meta's Llama 4 Community License requires companies with more than 700 million monthly active users to request a separate licence, requires "Built with Llama" to be displayed, and applies an acceptable use policy.
• Permissive licences: OpenAI gpt-oss-120b, Mistral Small 4, Google Gemma 4 and Qwen3-235B-A22B use Apache 2.0, while DeepSeek-V4-Pro uses MIT for both code and weights.
• Licences vary within a family: Mistral Large 4 uses the Mistral Ecosystem Open Weight License instead of Apache 2.0, the newer Qwen3.8-Flash-Next uses a Qwen community licence, and earlier Gemma versions use the Gemma Terms of Use.

Before using a model commercially, read the licence of the exact version you download rather than relying on the family name.