When an 'Open' AI Model Isn't Really Open Source

Every few weeks a lab announces that it is releasing a powerful model "openly," and the word does a lot of quiet work. When Mistral opened the weights to its trillion-parameter model this week, the coverage, ours included, reached for the language of openness. It is worth being precise about what that openness does and does not include, because the popular shorthand hides a real distinction.

Start with what "open weights" means. A trained model is, at bottom, an enormous grid of numbers, the weights, that encode what it learned. Releasing them lets anyone download the finished model, run it on their own hardware, fine-tune it and build products on top. That is a genuine and generous thing to do, and it is why a strong open-weight release is real news. You get the engine, fully assembled, free to use.

What you usually do not get is the recipe. Classic open source, in the sense the software world has meant for decades, implies access to the ingredients and the method: the training data, the code that did the training, and the freedom to reproduce the result from scratch. Almost no "open" frontier model offers that. The data in particular is nearly always withheld, both because it is commercially sensitive and because it is a thicket of copyright questions. So you can use the model, but you cannot see what went into it or rebuild it yourself.

The licence adds a second asterisk. Many open-weight releases come with conditions that a true open-source licence would not permit, such as caps on commercial scale or, as with Mistral's latest, a requirement that firms reselling the model as a service strike a separate deal. The weights are open. The terms are not unconditional.

None of this is a trick, and it is not a reason to be cynical. Open-weight models have genuinely widened access and loosened the grip of the handful of companies that can afford to train at the frontier. They also let outside researchers inspect how these systems actually behave. The point is only that "open" is a spectrum, not a switch. A released model can be free to run and still be a closed book about its own origins.

Why does the difference matter? Because the parts that stay hidden are the parts that bear on trust. Without the training data, you cannot fully audit what a model absorbed or where its knowledge came from, which is exactly what you would need to check it for bias or for memorised private information. So when a lab says it is being open, the useful question is not whether the claim is true but how far it goes. Open weights are a real gift. They are just not the whole workshop.