Overview
Fine-tune open-weight models on your own data and keep the weights.
The Nucleus AI API lets you take an open-weight base model, train it on your own data, and serve the result as a private model. Training runs on Nucleus GPUs, your data stays inside your account, and the finished weights can be exported and run anywhere.
Two ways in
There are two integration paths, and choosing the right one up front is the most important decision you will make.
Managed fine-tuning jobs
Upload a dataset, submit a job, poll until it finishes. Nucleus owns the training loop, the
hyperparameters and the retries; no ML engineering required. Start here unless you know you need
the loop. files → fine-tuning/jobs → samplers.
The low-level training loop
Drive gradients and optimiser steps yourself, one call at a time. Built for research teams doing
RL, DPO, distillation or custom losses: you bring the loop, Nucleus holds the GPUs.
training-runs → forward-backward → optim-step → checkpoints.
Base URLs
| Routes | Base URL |
|---|---|
| Platform and OpenAI-compatible | https://api.nucleus-ai.io/v1 |
| Anthropic-compatible | https://api.nucleus-ai.io/anthropic |
One API key works across both, sent as Authorization: Bearer nk_live_... on every route
except GET /v1/health.
Start building
Quickstart
Seven calls from a JSONL file to a running private model.
Authentication
Keys, scopes, rotation and what 401 and 403 responses look like.
Managed fine-tuning
The managed journey end to end: upload, train, evaluate, serve, export.
The training loop
Forward-backward passes, optimiser steps and checkpoints under your control.
API reference
Every endpoint, with parameters, scopes and example responses.