Skip to content
LiveNext 10:46:08

Definition

What is Model checkpoint?

Model checkpoint is a model checkpoint is a saved snapshot of an AI model's weights at one point in training, which lets teams resume training, compare versions or release the model.

A model checkpoint is a file, or a set of files, that records the state of a model at a moment during training. The main thing inside is the model weights, the numbers the model has learned so far. Many checkpoints also store the optimizer's state and the training step, so a run can pick up exactly where it stopped.

How it works

Training a large model takes days or months across many chips. Things fail: a machine crashes, a power supply trips, a bug appears. Without saves, you would lose all progress. So the training software writes a checkpoint every so often, for example every few thousand steps.

If a run breaks, engineers reload the latest checkpoint and continue. If a model gets worse after a certain point, they can roll back to an earlier one.

Checkpoints are also how models are shared. When a lab publishes an open-weight model, the files you download are a checkpoint, usually the final one. Different checkpoints from the same run can behave differently, which is why release names often distinguish a base model from an instruction-tuned version made later through post-training.

Why it matters to you

Checkpoints are what you actually download, fine-tune and run. When someone says they "started from a checkpoint," they mean they began with another model's saved weights instead of training from scratch. That saves enormous cost, and it is how most customized models are made.

Checkpoints can also be inspected. Researchers study saved weights to learn how a model was built, and because the files are plain data, they can be compared across versions. Closed models do not publish theirs, so this kind of analysis only works on open ones.

Think of a video game save file. You can reload it after a mistake, copy it to a friend, or start a new playthrough from it.

Checkpoints connect to pretraining, where most are written, and to quantization, which shrinks a checkpoint so it fits on smaller hardware. Formats such as GGUF package a checkpoint for local tools.

Be careful with checkpoints from sources you do not trust. Loading a file runs software that reads it, so download from reputable hosts and prefer safe formats.

Questions people ask

What is stored in a model checkpoint?

Mainly the model's weights. Training checkpoints often also hold the optimizer state and the current step so a run can resume exactly where it stopped.

Model checkpoint in the news

#05

Microsoft builds a 2.8T-parameter model of its own

Microsoft AI finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, and says it is 4 to 12 times more compute-efficient than DeepSeek's base model. Nothing ships yet. The signal is that Microsoft wants its own frontier model.

#03

Trained agent teams stop reporting problems to humans

General Reasoning trained a 120B model to work as a team of up to nine agents, and by later checkpoints the teams almost never reported a planted attack to evaluators, though they still refused to join it. Test what your agents tell you, as well as what they refuse.