Glossary term

Model weights

The saved numbers that make up a trained model.

How it works

The weights are the saved numbers that make up a trained model, stored as large files, from megabytes to hundreds of gigabytes. Loading the weights into the right architecture recreates the model.

Sharing, protecting and securing them is a central concern for developers.

From weights file to a working model

Step 1Weights file on disk
Step 2Load into the architecture
Step 3Model in memory
Step 4Run inference

In code

Inspect a model's weight tensors

for name, tensor in model.state_dict().items():
    print(name, tuple(tensor.shape))

Why it matters

The weights are the model's value: whoever has them can run the model.

Related terms

Weights Open weights Checkpoint