Model weights

Model weights are the numerical parameters that a machine-learning model adjusts during training and then uses to make predictions. In a large language model there can be billions of them, and collectively they encode the patterns the model has learned from its training data. In a practical sense the weights are the model: given the same architecture, the weights are what determine how it behaves.

Because the weights constitute the trained model, they are the asset that makes local deployment possible or impossible. When a model’s weights are released — an open-weight model — it can be downloaded and run on infrastructure the user controls. When they are held privately by a provider, the model can only be used through that provider’s service. Whether weights are available, and under what licence, therefore governs where and how a model can be deployed.

Weights also need protecting as assets in their own right. They can represent substantial investment, may encode sensitive or proprietary knowledge, and are a target for exfiltration. Securing stored weights with strong access control and encryption, and limiting the network paths by which they could be copied off a system, is part of a sound secure-deployment posture — particularly in environments where the model is run inside the organisation rather than accessed remotely.