Checkpoint

A checkpoint is a file that saves an entire training-in-progress model at a specific moment. Think of it like a save file in a video game — AI training can take weeks, and hardware failures are common along the way, so teams save progress periodically and, if something goes wrong, restart from the last checkpoint instead of from scratch.

When you hear news that a company "released a model," what actually happened is usually this: they let people download the final checkpoint produced after training finished. So when a team says they're also releasing "intermediate checkpoints," it means they're sharing not just the finished model but saved snapshots from along its training journey, for research purposes.

In the image-generation community, people often just call the model file itself a "checkpoint." That's also where the .ckpt file extension comes from.