model weights

Model weights are the numbers a neural network learns during training, often billions of them. Together they encode everything the model has picked up from its training data, and they are what actually produce its answers when it runs. In a real sense, the weights are the model.

This is why the phrase open-weight carries so much meaning. When a lab releases a model’s weights, it hands over the finished, runnable system, so anyone can use it on their own hardware. When the weights are kept private, you can only reach the model through the provider’s service.

Weights also explain why building frontier models is so expensive: producing a good set of weights requires vast amounts of data and compute, even though running the finished model afterward can be comparatively cheap.