Docs
Publishing models
How do I export or publish a model trained on Onrup?
Download the weights from the run, or publish to a model hub in one call. A model card is generated from the run — base model, dataset reference, hyperparameters, final metrics, GPU class and wall time. Publishing is never automatic and is not charged for.
The weights are yours
Every finished run exposes its trained weights for download through a signed URL. For an adapter that is a small file; for a merged model it is the whole thing.
This is the clearest test of whether a fine-tune belongs to you or to the platform that produced it, and it is worth applying to any vendor you evaluate.
Record the pairing
An adapter is only usable against the exact base model and version it was trained on. Against a different version it produces degraded or nonsensical output, and nothing in the file format prevents the attempt.
The generated model card records the pairing, which is the main reason to keep the card alongside the weights rather than treating it as documentation.
Publishing is a call you make
Nothing is published automatically. A publication pushes weights and the model card to a public or private hub repository, with optional inclusion of checkpoints and metrics. Datasets are never included.
A run can have many publications — republishing creates a new one rather than mutating the old. Publishing is not charged for.
The licence follows the weights
A fine-tune inherits the base model’s licence. Publishing a derivative of a model with a restrictive licence does not relax those terms, and the obligations pass to whoever downloads it.
The generated card records which licence applies. Where a base model carries a naming requirement or a revenue threshold, that is a term to read before publishing rather than after.
Frequently asked questions
Can I publish privately?
Yes, and it is a sensible default. A private repository is a portability mechanism: the model exists somewhere we do not control, without being public.
What is in the generated model card?
Base model, dataset reference, template and version, hyperparameters, final loss and throughput, GPU class, wall time and licence. Intended use and limitations are left for you, because those are judgements rather than facts about the run.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.