Model weights are the numerical parameters a machine learning model learns during training. In a neural network they are the strengths of the connections between units, adjusted again and again until the model's outputs match the training data well. A trained large language model may have billions of weights stored in large files; anyone with those files and suitable hardware can run the model.
Weights are costly to produce, since training needs large amounts of data and compute, but cheap to copy. That makes them awkward as assets. Their legal protection as intellectual property is unsettled in many places, they are usually guarded by trade secrecy and licences rather than ownership registers, and their value can fall quickly when a better model appears. Proposals to tokenise model weights face the same problems: a token can record rights or revenue shares, but cannot stop a copied file from being used, so such schemes remain rare.
Ref 1300000379