Training on Baseten

Train models on managed GPUs with Loops or Training Jobs and deploy any checkpoint to production inference.

Baseten trains models on managed GPUs and deploys the resulting checkpoints to production inference on the same platform. Train directly with Loops, a Python SDK for LoRA fine-tuning and RL on a curated set of base models, or bring your own container and training code with Training Jobs.

Choose your path

• Fine-tuning or running RL on a supported base model: Loops provisions a dedicated trainer and paired sampler, and each training step is an API call from a Python loop you write. You bring the data; Baseten runs the infrastructure. Start with the Loops overview.
• Already training elsewhere: Your Axolotl config, TRL script, or custom loop runs unchanged in a container. Baseten provisions the GPUs, syncs checkpoints as your job saves them, and deploys any checkpoint as a production endpoint. That’s Training Jobs.

Training Jobs runs your program to completion on hardware you choose. With Loops, your program calls a trainer and sampler through the API.

	Training Jobs	Loops
Training code	Any container image	A Python training loop (baseten-loops SDK)
Models	Any	Supported base models only
Hardware	You declare GPUs in a Truss config	Baseten picks GPUs for the base model
Inference path	Deploy any synced checkpoint with one CLI command	Sampler serves new weights live during training; deploy checkpoints when ready
Lifecycle	Job runs to completion and exits	Session stays live until you run baseten loops run deactivate
Availability	All workspaces	All workspaces
Documentation	Jobs docs	Loops docs

Next steps

Loops quickstart

Train a step, sample from the tuned weights, and deploy a checkpoint.

Jobs quickstart

Run your first training job and deploy the result.

ML Cookbook

Production-ready examples for frameworks and models.