Export any checkpoint, any time. Run it on hardware you control.
Pick a checkpoint from any run. Choose a LoRA adapter, merged safetensors or GGUF. Send it to a signed URL or your Hugging Face repository, and load it into vLLM, transformers or llama.cpp.
POST /v1/checkpoints/ckpt_000675/export
{ "format": "merged-safetensors", "destination": "signed-url" }
202 Accepted
{ "operation": "op_91c4", "format": "merged-safetensors", "status": "running" }Any checkpoint of any run
Every checkpoint stays in your account until it expires or you delete it, and any of them exports.
Three formats
A LoRA adapter, merged safetensors or GGUF, chosen for where the weights will run.
Two destinations
A time-limited signed URL, or a push to the Hugging Face repository you name.
Runs anywhere, offline included
The export loads into vLLM, transformers or llama.cpp, including fully offline and on-premises deployments.
From a checkpoint to a file you can load
Three choices and one call. The export runs in the background, and the weights arrive where you sent them.
A training run writes checkpoints as it goes, and every one of them stays in your account until it expires or you delete it. List them across all runs, newest first, or resolve one from its nucleus:// path. Export the newest, or the one your evaluations scored best. Clear the expiry on the ones you serve from, so they stay.
Results from LLMs powered by Nucleus
downtime prevented this month · Predictive Maintenance · Manufacturing
from site reading to client report · Real-Time Field Data Collection · Field engineering
yield shortfall traced to one cause · Agriculture Yield Analysis · Agriculture
Weight export, in four columns
Checkpoints
- Sampler checkpoints hold servable weights
- Training checkpoints hold resumable state
- Kept until they expire or you delete them
- ttl_seconds set at creation, changed later
- Listed across all runs, newest first
- Resolved from a nucleus:// path
Formats
- lora-adapter, for your own serving stack
- merged-safetensors, for vLLM and transformers
- gguf, for llama.cpp
- One format field per export call
Destinations
- signed-url, a time-limited download URL
- hf-hub, a push to the repository in hf_repo
- The raw archive by signed URL, for backup and transfer between Nucleus environments
The call
- POST .../checkpoints/{checkpoint_id}/export
- The train scope
- 202 Accepted with an operation id
- Poll /v1/operations/{id}, or register a webhook
- Idempotency-Key on automatic retries
Pay for what you keep
Weight export runs on the storage meter. Datasets, checkpoints and exported weights are metered by the GB-hour, so you only pay while you keep them.
The rest of Own
Checkpoints
Kept until you delete them
Data residency
Choose your region
Bring your own cloud
The platform in your account
Storage
Metered by the GB-hour
Read on
Serving and exporting
Take finished weights to production by serving them on Nucleus or exporting them to run anywhere.
Checkpoints
Save, inspect, export and manage the weights produced by training runs.
Platform
Account limits, long-running operations, webhooks and the health probe.
Take the weights with you.
Any checkpoint, any time, in a format you can load anywhere.