Skip to content
BENIBIX Learn

Model Weights File

AI💻 Running LocallyLocal Models

Also known as Model Weights, Model Checkpoint, Weights File

The file that holds all the numbers a model learned.

Think of it like

Think of it like a recipe book. The cook is the program, and the book holds all it learned.

Example

Jack downloads an open model from a model hub. The weights come as four files of about 5 GB each, plus a small config file.

Why it matters

The file size tells you at once whether a model will fit on your machine.

Definition

A model weights file holds the numbers a model learned in training, its weights, saved to disk. It is the model itself: the program that runs it is the same for many models, and the weights are what make each one different. Its size is about the number of weights times the bytes each takes. An 8 billion weight model at 16 bits, which is 2 bytes each, is about 16 GB. Common formats include safetensors and GGUF.
A weights file is a long list of numbers. Its size is the count times the bytes each takes.

Common mistake

Opening weights files from unknown sources in old formats. Some old formats can run hidden code when loaded, so prefer safe ones.

Key takeaways

  1. 1
    The learned numbers. Saved to disk.
  2. 2
    Size from count. Weights times bytes.
  3. 3
    Format matters. Some are safer.