Every run's checkpoints, kept until you delete them
A training run writes a checkpoint whenever you ask for one. Each one stays in your account beside the dataset that trained it, for as long as you set. Compare them, promote one to an endpoint, and delete the rest.
Two types, one path
A training checkpoint holds the full state needed to resume. A sampler checkpoint holds servable weights. Each carries a nucleus:// path that samplers and Load state accept.
Retention you set
Pass ttl_seconds when you create a checkpoint, change it later, or pass null to keep it until you delete it.
Kept beside the data
The dataset a run trained on stays in your account with its checkpoints and trains nobody else's model.
Paid for while kept
Storage is metered by the GB-hour at $0.10 per GB a month, so a checkpoint stops costing when it goes.
What happens to your runs
Four things the platform does with a run's checkpoints, in the order you meet them.
A run writes a checkpoint whenever you ask for one, under a name you choose. Training checkpoints carry everything needed to resume; sampler checkpoints carry weights a sampler can serve. The list shows each one with its type, size and expiry, beside the dataset the run trained on.
Results from LLMs powered by Nucleus
downtime prevented this month · Predictive maintenance
daily fuel saved across the fleet · Fleet optimisation
monthly revenue misattributed · Ad performance tracking
yield shortfall traced to one cause · Agriculture yield analysis
What a checkpoint can do
Types and paths
- Training checkpoints, resumable with optimiser state
- Sampler checkpoints, servable weights
- A nucleus:// path on every checkpoint
- Stable names, overwrite on training checkpoints
- Idempotency keys on create and export
Retention
- ttl_seconds at creation
- Expiry changed later, counted from the call
- null keeps a checkpoint until you delete it
- Permanent deletion by run and checkpoint id
Access
- Private by default
- Publish and unpublish with the admin scope
- Time-limited signed URL for the raw archive
- Another account's published weights through a weights access token
Use
- Create a sampler from the path
- The model on the OpenAI and Anthropic routes
- Load state to roll back mid-loop
- Evaluate against a baseline with compare_to
- Export as LoRA adapter, merged safetensors or GGUF
Metered while you keep it
Storage
Datasets, checkpoints and exported weights. Metered by the GB-hour, so you only pay while you keep them. The usage API reports storage GB-hours for any date range, grouped by day, project or base model.
Pick your path
Weight export
Any checkpoint, any format
Data residency
Choose your region
Bring your own cloud
The platform in your account
Storage
Metered by the GB-hour
Read on
API reference · Checkpoints
Save, inspect, export and manage the weights produced by training runs.
Guide · Serving and exporting
Take finished weights to production by serving them on Nucleus or exporting them to run anywhere.
API reference · Evaluations
Score a model against a dataset and compare it with a baseline.
Keep every checkpoint for as long as you choose
Retention you set, a path you can serve from, and deletion that is yours to call.