AI 101

Checkpoint

Also known as: model checkpoint · weights file

Basics

A checkpoint is a saved snapshot of a model’s weights at a particular moment. It’s how models get distributed — the file you download is a checkpoint (or a quantized version of one).

  • During training: checkpoints let you resume if training crashes, or pick the best intermediate version.
  • In release: a “7B checkpoint” is the raw weights file; tooling like Ollama serves it.
  • Fine-tuning: you start from a base checkpoint and adjust from there.

Practical: when a page says "download the 7B", it means the checkpoint. The size in GB depends on precision — see [quantization](/glossary/quantization/).