FAQ
Frequently asked questions
Including the awkward ones?
Yes. Several answers below are "no" — we do not run on your hardware, we do not accept custom training code, we do not do multi-node distribution, and we do not have a compliance certification to point at today. An FAQ that only asks flattering questions is a brochure.
What Onrup is
What is Onrup?
An API-first platform for fine-tuning open-weight language models and serving them. You upload a dataset, pick a base model and objective, run the training, gate the result against the model it would replace, and deploy it to an OpenAI-compatible endpoint. Compute is serverless across multiple providers.
Who is it for?
Teams who have a specific task, some data about how that task should be done, and either a cost problem or a portability requirement. It is not a general model API, and it is not a research platform for custom training code.
What makes it different from other fine-tuning platforms?
Three things: cost is authorised and reserved before compute is leased rather than measured afterwards; a blinded evaluation gate sits between training and deployment and can block it; and every model is downloadable or publishable, so leaving costs you a base URL change.
Is there a free tier?
Yes. Account creation, dataset upload, validation reports and the full catalogue are free and need no card. The free tier deliberately does not include compute — the point is to let you find out whether your data is usable before paying anyone anything.
Models and training
Which models can I fine-tune?
43 open-weight models across 13 families, from 350M to 117B parameters, covering the sizes and objectives most fine-tunes actually need. The catalogue is curated rather than open, which is the main thing that can block a migration — check your base model is on it before planning one.
Can I bring my own training code?
No. Runs execute within validated templates, which is what makes cost forecasting, resumption and the evaluation gate possible. If you need an arbitrary training loop, a general serverless compute platform is the honest recommendation and we will say so.
Which objectives are supported?
Supervised fine-tuning, preference tuning and reinforcement learning from verifiable rewards. Not every model supports all three — preference tuning needs memory for a reference model, which puts it out of reach at the largest sizes.
How long does a fine-tune take?
A LoRA run on an 8B model over a few thousand examples is typically well under an hour of GPU time. Wall time also includes waiting for capacity in your chosen class, which the forecast indicates before you commit.
Can I train a model that does not fit on one GPU?
Four-bit training brings a 70B model inside a 48GB class and the largest models in the catalogue inside 80GB. Beyond that we do not do multi-node distribution, and a platform built around it is the right answer.
Cost
How is compute priced?
Per GPU-second above a one-minute floor, at a published rate per GPU class. Training and serving are priced separately because they are different workloads — training is bursty and bounded, serving is resident and continuous.
Will I be surprised by a bill?
The system is built specifically so that you are not. Every job is costed and reserved against your spend limit before compute is leased, counting open reservations as well as settled spend. A job that would breach the ceiling does not start.
Do I pay while waiting for capacity?
No. Metering starts when your job is on the GPU. Queue time and teardown are not billed.
What does an idle endpoint cost?
Nothing, if it is configured to scale to zero. An always-warm endpoint bills continuously for the time it is resident, whether or not requests arrive — which is the honest description of reserved capacity.
Is Onrup cheaper than the alternatives?
On directly comparable dedicated H100 capacity, yes — our serving rate is $2.99 per GPU-hour against published rates of $5.49 to $7.00 elsewhere. On per-token fine-tuning there is no like-for-like comparison, because we bill by GPU-second rather than by token, and which is cheaper depends on your dataset and traffic shape.
Compute and infrastructure
Where does the compute come from?
Serverless compute across multiple providers. You choose a GPU class and a scaling mode; capacity is found, leased, monitored and torn down. Which provider fulfils a given job is not surfaced, and sourcing across several rather than one is what keeps availability from tracking any single vendor’s inventory.
Can I run Onrup on my own hardware?
No. The control plane is managed and compute is serverless. If the model must run inside your own network, a vendor that ships software rather than a service is the right answer.
Can I choose a region?
Not today. You choose a GPU class, not a location. If data residency is a contractual requirement, say so before starting rather than after.
What happens if capacity is not available?
The run stays queued and the API returns a wait indication rather than failing silently. For serving, a request to a cold endpoint queues within bounds you set, and is refused with a retry hint rather than queueing indefinitely.
Data and ownership
Do I own the trained model?
Yes. Weights are downloadable from every finished run, and publishing to a model hub is one call. This is the clearest test of whether a fine-tune belongs to you, and it is worth applying to every vendor you evaluate.
Is my data used to train anything else?
No. Datasets stay in your tenant’s storage and are used only for the runs you start. Publishing pushes weights and a model card, never data.
What dataset formats do you accept?
Conversation, instruction, preference and prompt-only shapes, as JSONL, CSV or Parquet. A format adapter converts whichever you have into the trainer-ready form, so the dataset stays independent of the model choice.
Can I delete my data?
Yes, including derived artefacts. Deletion is recorded in the audit log, which is append-only.
Evaluation and deployment
What is the evaluation gate?
A blocking check between training and deployment. The candidate and the model it would replace answer the same held-out cases under randomised aliases, a judge scores them blinded, and the result is bootstrapped for a confidence interval. The verdict is pass, fail or inconclusive, and only pass unlocks a deploy or publish.
Can I override a failed gate?
A person can approve a new evaluation suite. What they cannot do is mark a recorded failure as passed. A gate somebody can wave through at the end of a long week is a dashboard, not a gate.
Does the endpoint work with my existing OpenAI client?
Yes. Change the base URL and the key. Chat completions, streaming and tool calling all use the same request and response shapes.
Can I serve several fine-tunes economically?
Yes, if they share a base model and version. Adapters are co-resident against one resident base, so the marginal cost of an additional variant is megabytes rather than gigabytes.
Something not answered?
The documentation covers the API in full, the glossary defines the vocabulary, and contact reaches a person.
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.