noun
A checkpoint is a snapshot of an AI model’s state, including its weights and training progress, saved at regular intervals or key milestones. This allows developers to resume training from that point if interrupted or to compare performance across different stages.
Checkpoints are crucial in large-scale AI training, where interruptions (e.g., hardware failures or budget limits) can waste weeks of work. They also help fine-tune models by letting engineers revert to a stable version before a problematic update.
For developers, checkpoints reduce costs by avoiding redundant training and enable experimentation without risking the entire model. For example, training a model like Stable Diffusion V2 typically costs $500–$1,000 per week on cloud GPUs; checkpoints let teams pause and resume without losing progress.
After a server crash, the team restored the AI model from a checkpoint saved 3 days earlier, saving 2 weeks of retraining.
In other languages
See also: model fine-tuning