Checkpoint
A saved snapshot of a model’s parameters at a moment in training, so you can stop, resume, or ship it.
Think of it like
A save point in a video game — freeze progress now, come back to exactly here later.
Example
If a training run crashes at hour 300, you reload the last checkpoint instead of starting over.
How it actually works
A checkpoint is the full set of weights written to disk mid- or post-training. It lets you resume interrupted runs, compare versions, roll back a bad update, or deploy a specific state. “The model you use” is really one chosen checkpoint.
For product teams
Every deployed model is a specific checkpoint — pin the version so behavior doesn’t shift under you.
For engineers
Serialized parameter (and optimizer) state at a training step, used for resumption and deployment.
Related
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome