What Is a Checkpoint in AI Model Training?
Training a frontier model can take months — checkpoints are the safety net (and the comparison tool) that makes a run like that manageable.
A saved snapshot mid-training
A checkpoint is a saved snapshot of a model's parameters at a specific point during training — not the final model, but a complete, usable copy of it as it existed after some number of training steps. Training regularly saves new checkpoints rather than only keeping the very last state.
Why this matters for runs that take months
A frontier training run can take weeks or months on enormous compute clusters, and hardware failures or bad data batches do happen. Checkpoints let training resume from the last good save instead of starting over from scratch if something goes wrong partway through — the difference between losing hours and losing months.
Checkpoints as a comparison tool
Because each checkpoint is a fully usable model, teams can evaluate several checkpoints from different points in training against each other, to see how capability progressed over the run and pick the point where the model was strongest before quality started to plateau or (via overfitting) begin to degrade.
Frequently Asked Questions
What is a checkpoint in AI training?
A saved snapshot of a model's parameters at a specific point during training — a complete, usable copy of the model as it existed after some number of training steps, not just the final result.
Why do training runs save checkpoints regularly?
So training can resume from the last good save if something goes wrong (hardware failure, bad data), instead of restarting from scratch on a run that can take weeks or months.