What does 'open weights' mean, and is it the same as open source?
No. Open weights means the trained model parameters are downloadable; open source, in its established meaning, requires considerably more — and the distinction is actively contested because a great deal of marketing depends on blurring it.
What open weights provides. The weights — the numerical parameters resulting from training — can be downloaded, run locally, fine-tuned and deployed. That is genuinely valuable: it permits offline use, private processing, modification, and independence from any provider.
What is usually not provided:
The training data, or even a full description of it. Without it, the model cannot be reproduced, and its behaviour cannot be fully explained or audited.
The training code and procedure in reproducible detail.
The cost. Reproducing a large model requires computation most organisations cannot access, so even complete disclosure would not make it practically reproducible.
Why the open source comparison breaks down. For software, source code is the preferred form for making modifications — with it, anyone can rebuild and change the program. For a model, the weights are closer to a compiled binary than to source: you can modify them at the margins, but you cannot rebuild them from scratch or understand how they came to be as they are.
The licence question. Several prominent "open" models carry licences with restrictions genuinely incompatible with the Open Source Definition — usage limits above a scale threshold, restrictions on competing uses, or acceptable use policies constraining applications. A licence that restricts who may use it or for what is not an open source licence, by the accepted definition.
The Open Source Initiative has published a definition of open source AI requiring, among other things, sufficient information about training data to permit a substantially equivalent system to be built — and several widely described "open source" models do not meet it.
Why it matters practically: check the actual licence before building on a model, since restrictions may bite at exactly the point a product succeeds; and understand that auditing a model for bias or contamination is impossible without knowing what it was trained on.