Definition
A model checkpoint is a saved state from a particular point in training. A weights-only checkpoint may be enough for inference. A resumable training checkpoint usually also needs optimizer state, learning-rate scheduler state, the current step or epoch, mixed-precision state, and enough random-number and data-loader state to continue with the intended training trajectory.
The file is only one part of reproducibility. Restoring the same run may also require the model code, tokenizer, configuration, dependency versions, data version, and distributed-training layout. Even then, nondeterministic hardware operations can prevent a bit-for-bit continuation.
PyTorch's saving and loading guidance makes the practical distinction explicit: learned parameters live in a model state dictionary, optimizer objects have separate state, and a general checkpoint stores both when training must resume. A checkpoint intended for distribution should be treated as a model artifact, with provenance, integrity checks, format controls, and compatibility metadata.
Distinguish it from nearby terms
A model checkpoint stores learned and training state. An application release also includes code, configuration, migrations, and deployment metadata. An agent recovery checkpoint stores workflow state such as completed actions, tool results, and pending work; model weights may not change at all.
Check your understanding
You restore the model weights and epoch number, but not Adam's moment estimates or the learning-rate scheduler. Can the run continue, and in what sense is it no longer the same training trajectory?