Checkpoint
A saved snapshot of weights, and usually optimiser state and step count too, so a run can be resumed or an earlier model recovered.
What people get wrong
Weights alone are enough to run a model but not to resume training it. Without optimiser state the momentum and variance estimates restart from nothing, and the loss visibly jumps.
A definition is not an understanding. This is taught properly in Checkpoints and reproducibility, part of PyTorch.
Learn this properly