Overview
Dataset in, endpoint out.
What does the workflow look like?
Upload a dataset and read its validation report, forecast the run, submit it — which is where cost is authorised against your limit — watch it train, gate the result against the model it would replace, then deploy or publish. Five calls, with the gate in front of the expensive step.
- 01
Upload and validate
A resumable multipart upload sends bytes straight to storage. Completing it triggers validation, which reports record counts, token-length percentiles, duplicate rate and how many examples would be truncated. Nothing has cost anything yet.
- 02
Forecast
Given a model, a template and a dataset size, the forecast returns an estimated cost and an indication of how long you may wait for capacity in your chosen class.
- 03
Submit
Submission estimates the maximum cost, reserves it against your spend limit counting open reservations, and queues the run. A job that would breach the ceiling is refused here, before any compute is leased.
- 04
Train
The run moves through matching, provisioning, training, checkpointing and uploading, emitting resumable events throughout. Checkpoints mean an interruption costs the time since the last one rather than the whole run.
- 05
Gate
The finished candidate is scored blinded against the model it would replace, on held-out cases from your own data. The verdict is pass, fail or inconclusive, and only pass unlocks the next step.
- 06
Deploy or publish
Create an OpenAI-compatible endpoint, always-warm or scale-to-zero. Or download the weights, or push them to a model hub with a generated model card — or all three.
What you choose, and what you do not
| You decide | Handled for you |
|---|---|
| Base model, from 43 in the catalogue | Which provider fulfils the job |
| Objective — supervised, preference or reinforcement learning | Environment setup and dependency pinning |
| GPU class, from 13 | Acquiring, monitoring and tearing down capacity |
| Spend limit and the cost ceiling per job | Checkpointing, resumption and cleanup |
| Scaling mode — always warm or scale to zero | Metering, reconciliation and usage delivery |
| Which cases the gate runs against | Blinding, alias randomisation and bootstrapping |
The right-hand column is undifferentiated work. Somebody has to do it, and doing it yourself does not make your model better.
What it will not do
- Run arbitrary training code. Runs execute within validated templates, which is what makes forecasting, resumption and the gate possible.
- Distribute a run across multiple nodes. Four-bit training brings the largest models in the catalogue inside a single 80GB class; beyond that, a distributed platform is the honest answer.
- Run inside your own cloud account or on your own hardware.
- Serve image, video or audio models. The catalogue is language models only.
Frequently asked questions
How many API calls is that?
Five for the common path: upload, submit, gate, deploy, and a read of the validation report in between. Everything else is optional.
Do I have to use the console?
No. The API is the contract and the console is a client of it — there is nothing the console can do that the API cannot.
Can I skip the gate?
The gate is what makes a deploy or a publish possible, so no. It is also the cheapest part of the sequence, and it is the step that answers the question the rest of the work was for.
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.