Reference · Glossary

Model weights

Last updated

The **learned numbers** inside a neural network that encode patterns from training — billions of parameters saved as files.

When to use

Discussing training cost, fine-tuning, downloading open models, quantization, or why inference reads weights but doesn't rewrite them.

When not to

End users rarely touch weights directly — they call an API or app that loads weights on a server or GPU.

Example

Llama weights on disk might be 4–70 GB; inference loads them once, then runs forward passes for each user prompt.