What is the difference between open-weight and open-source AI?

You will often hear models like Llama or DeepSeek described as open source. Strictly speaking, many of them are better described as open-weight. The difference sounds technical, but it matters if you plan to build products or do research.

An AI model has several ingredients. There is the training data, the huge collection of text, code or images it learned from. There is the training code, the software that turned data into a model. And there are the weights, the billions of numbers that make up the finished trained model.

An open-weight model releases the weights, so you can download and run the finished model, fine-tune it on your own data and build applications with it. But the training data and often the full training process stay private. You can use the model, but you cannot fully reproduce how it was made.

A truly open-source model, in the traditional software sense, would release everything needed to understand and rebuild it, including training code, detailed information about the data and the weights, under a licence that allows anyone to use, modify and share it for any purpose. The Open Source Initiative has published a formal definition of open-source AI along these lines. Only a small number of models, mostly from research groups, meet that full standard.

Why does it matter?

1. Licence terms differ. Some open-weight models use very permissive licences such as Apache 2.0 or MIT. Others include restrictions, such as acceptable-use rules, limits for companies above a certain number of users, or requirements to credit the model. Read the licence before using a model commercially.
2. Transparency differs. Without the training data, it is harder to check for bias, copyrighted material or why a model behaves a certain way.
3. Reproducibility differs. Researchers can study an open-weight model's behaviour but cannot retrain it from scratch to test changes.

For most practical uses, such as running a model privately or fine-tuning it for a business task, open-weight is perfectly good. The key point is not to assume "open" means "no rules." Check the licence, check what was actually released, and choose a model whose terms fit what you plan to do.