What are model weights?
Model weights are the numbers a model learned during training, stored so the model can be run again. Strictly, weights are the multipliers on connections between units and biases are the offsets, but in everyday use “the weights” means all of a model’s learned parameters. Weights are useless without the matching model architecture, which says how the numbers are arranged.
Weights are distributed as files, commonly in the safetensors format, and run to tens or hundreds of gigabytes for a large model. A checkpoint is a set of weights saved partway through training.
Why model weights matter
Whoever holds the weights controls the model. An open-weight model publishes its weights under a license, so a company can download them, run them on its own hardware, fine-tune them, and keep that exact version indefinitely. A closed model exposes only an application programming interface (API), and the vendor can change or retire it at will. For a regional bank that needs its document models to behave the same way at the next audit as at the last one, holding the weights makes that possible. Kirk Tech Solutions typically keeps a client’s fine-tuned weights in the client’s own model registry as a company asset.
What to watch for
- Read the license, because “open” ranges from fully permissive to restrictions on commercial use.
- Download weights only from the publisher’s official source, and prefer safetensors over formats that can execute code when loaded.
- Record the exact weight version running in production, since two releases with the same name can behave differently.