Pillar

Your weights, your endpoint.

Do I own a model I fine-tune on Onrup?

Yes. Every finished run exposes its trained weights for download, and publishing to a model hub is one call with a generated model card. Serving implements the OpenAI chat-completions API, so moving your application elsewhere is a base URL and a key.

The test worth applying to every vendor

Can you download the weights? It is the clearest single question in this market, because it separates a platform that produced an asset for you from one that produced a capability you rent. A platform that trains a model and will not export it has given you something that cannot be run elsewhere, audited independently, or kept if the relationship ends.

Of the platforms we compare ourselves against, 2 cannot export a customised model at all. That is not a criticism of their products — it is a structural property of building on closed weights, and for some teams it is an acceptable trade. It should be a decision made at the start rather than discovered later.

Portability is three artefacts, not one

People think portability means having the weights. It means having the weights, knowing exactly which base model and version they apply to, and still holding the dataset that produced them.

The pairing is the part that gets lost. An adapter applied to a different version of its base produces degraded or nonsensical output, and nothing in the file format warns you. That is why the generated model card records base model, version, template, hyperparameters and final metrics — it travels with the weights rather than living in a wiki page that gets archived.

How to keep a model portable, and test it →

Publishing is a call you make

curl https://api.onrup.com/v1/runs/$RUN_ID/publications \
  -H "Authorization: Bearer $ONRUP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "repo_id": "your-org/support-triage-v3",
    "private": true,
    "include_metrics": true
  }'

Nothing is published automatically and publishing is not charged for. A private repository is a portability mechanism as much as a distribution one: the model exists somewhere we do not control, without being public.

The licence follows the weights

A fine-tune inherits its base model’s licence, and those obligations pass to anyone you give the model to. Publishing a derivative of a model with a restrictive licence does not relax those terms.

The catalogue states the licence for every base model, and the generated card records which one applies to your derivative. The clauses that surprise people are naming requirements, restrictions on using outputs to train other models, and revenue thresholds that convert a free licence into a commercial negotiation as a company grows.

Licence terms for every base model →

Frequently asked questions

Can I download my fine-tuned model?

Yes, from every finished run, through a signed URL. For an adapter that is a small file; for a merged model it is the whole thing.

What do I need to keep for the model to be usable elsewhere?

Three things: the weights, the exact base model and version they were trained against, and the dataset. An adapter without its base pairing is close to useless, and the dataset is the insurance that lets you rebuild if anything else is lost.

Does publishing cost anything?

No. Publishing to a model hub is not charged for, and it is never automatic — nothing leaves your tenant until you make the call.

Is my training data published along with the model?

No. A publication pushes weights and a model card. Datasets are never included.

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.

Get your API key Read the docs