Glossary

Model weights

The numerical parameters a model learns during training — the artifact that constitutes the trained model and is deployed, shared, or sold.

Weights are what training produces and what inference runs; releasing them publicly (open weights) hands anyone the full model. Memorization research shows models can sometimes reproduce training examples, which is why data licences treat weights as a derivative of the data rather than something unrelated to it. Terms about keeping, transferring, or open-sourcing weights are where a data deal continues to bind after training ends.

Why it matters

For a data owner, what the licensee may do with the weights matters as much as what they may do with the files.

See also

← All terms