Open weights vs open source

"Open weights" means you can download and run a model. "Open source" means you also get the training code, data documentation, and process needed to reproduce or audit it. Most models marketed as "open" are open weights only.

When a lab releases a model's weights (the learned parameters that define its behavior), you can run it on your own hardware, fine-tune it for your use case, and integrate it into products, subject to the license terms. Open source AI goes further by providing the full training recipe: the code, data documentation, and methodology needed to reproduce the model from scratch or audit how it was built. The Open Source Initiative formalized this distinction in 2024. Meta's Llama models, for example, are open weights with a custom license that restricts certain commercial uses, making them meaningfully different from a fully open source release.

Builder example

This distinction has direct legal and operational consequences. If you plan to self-host, fine-tune, redistribute, or build a commercial product on top of a model, the license terms determine what you can and cannot do. Some open-weights licenses prohibit use above a revenue threshold, restrict certain industries, or require attribution. Assuming "open" means "free to do anything" can create legal exposure.

Common confusion: "Open" is a marketing term with no single definition. A model can be open weights while having restrictive commercial terms, making it less permissive than its branding suggests.