Fine-tuning and inference control plane
Build models, not infrastructure.
Fine-tune 43 open-weight models, gate every candidate against its baseline, and serve the winner on your own OpenAI-compatible endpoint — on serverless compute across multiple providers, from $0.09 an hour.
A price you authorise before it is spent
Every run is costed and authorised against your spend limit before any compute is leased. Not a post-hoc invoice, not a credit alarm after the fact — a hard gate in front of the work.
Your weights, your endpoint
Download the adapter or the merged weights whenever you like, or publish to Hugging Face in one call with a generated model card. Serving speaks the OpenAI chat-completions API, so leaving costs you a base URL change.
It ships only if it wins
Every candidate is scored against the model it would replace, blinded, on your own cases. The gate returns pass, fail or inconclusive — and only pass unlocks a deploy or a publish.
Serverless compute across multiple providers
You choose a GPU class and a scaling mode. Capacity is found, leased, watched and torn down across multiple providers, so an idle endpoint costs nothing and a busy one never queues on a single vendor.
The argument
The hard part isn’t training. It’s knowing it worked.
Fine-tuning an open-weight model is a solved problem in the narrow sense: the recipes are published, the libraries are mature, and the hardware is rentable by the hour. What is not solved is everything around it — a run that costs three times what you expected, a model that scores well and disappoints in production, and weights you cannot get at when you want to leave.
So the platform is built around three refusals. It will not lease compute before the cost is authorised against a limit you set. It will not deploy a model that has not beaten the one it would replace on your own held-out cases, blinded. And it will not hold your weights hostage — every finished run is downloadable, and publishing to a model hub is one call.
Underneath, compute is serverless across multiple providers. You pick a GPU class and a scaling mode; capacity is found, leased, watched and torn down. An idle endpoint costs nothing and a busy one is not waiting on any single vendor’s inventory.
How it works
Five calls, with the gate in front of the expensive one.
- 01
Upload and validate
Resumable upload straight to storage. The validation report gives token-length percentiles, duplicate rate and truncation risk before you spend anything.
- 02
Forecast and submit
The run is costed and reserved against your spend limit. A job that would breach the ceiling does not start — it is a gate, not an alarm.
- 03
Train and watch
Progress streams as resumable events. Checkpoints mean an interruption costs the time since the last one, not the whole run.
- 04
Gate the candidate
Blinded against the model it would replace, on your held-out cases. Pass, fail or inconclusive — and only pass unlocks a deploy.
- 05
Serve or publish
An OpenAI-compatible endpoint, always-warm or scale-to-zero. Or download the weights and take them somewhere else entirely.
Price
A published rate card, per GPU class.
Compute is metered per GPU-second above a one-minute floor. Training and serving are priced separately because they are different workloads. On directly comparable H100 capacity, our serving rate is 2.3× lower than the dearest published rate in this market.
Reference
Everything you need to decide, before you sign up.
Model catalogue
43 base models with memory thresholds, supported objectives and licence terms.
Glossary
124 terms from LoRA to reward hacking, defined without marketing.
Guides
Whether to fine-tune at all, how much VRAM you need, and why fine-tuned models fail.
Comparisons
Head to head against every serious alternative, including where each one wins.
Use cases
Triage, extraction, agents, moderation — with the dataset shape and the pitfall for each.
Cost calculator
What a run costs, from model, dataset size and GPU class.
Common questions
What is Onrup?
An API-first platform for fine-tuning open-weight language models and serving them on an OpenAI-compatible endpoint, on serverless compute across multiple providers. Cost is authorised before compute is leased, and a blinded evaluation gate sits between training and deployment.
How much does fine-tuning cost?
Compute is metered per GPU-second above a one-minute floor, from $0.09 per GPU-hour on the entry class. A LoRA run on an 8B model over a few thousand examples is typically well under an hour of GPU time.
Do I own the model I train?
Yes. Weights are downloadable from every finished run, and publishing to a model hub is a single call with a generated model card. Serving speaks the OpenAI chat-completions API, so moving elsewhere is a base URL change.
Which models can I fine-tune?
43 open-weight models across 12 families, from 350M to 117B parameters, supporting supervised fine-tuning, preference tuning and reinforcement learning depending on the model.
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.