Fine-tuning
AvailableTeach an open model your task, then serve it.
Upload a training set, pick a base model, and get an adapter you can serve on a dedicated endpoint, with loss curves and GPU telemetry while it runs.
- Pricing
- Billed per GPU-hour
- Layer
- Training
- Train on
- Trainable catalog models, Versioned datasets, Uploaded training sets, Pipelines, Dedicated endpoints, Managed deployments
Why Fine-tuning
Change how a model behaves, not only what it knows.
A LoRA fine-tune trains a small adapter on top of an open base model, so you change its behavior on your task without paying for a full-weight run. Point a job at a dataset and a base model, watch it train, and promote the result when the evals clear.
LoRA adapters
Train on Qwen3 and GLM bases without full-weight cost.
Live metrics
Loss, throughput and GPU telemetry as the job runs.
One-click serving
Deploy the result to a dedicated endpoint.
How it works
LoRA fine-tunes on the open models in our catalog.
In the console
Everything in one place.
Fine-tuning lives in the LegionEdge cloud console, next to every other service in the workspace, under one account and one identity.
Pricing
Billed per GPU-hour
Fine-tuning is billed per GPU-hour, at compute rates, for the GPUs a job uses. A LoRA run trains an adapter, not the full weights, so you avoid full-weight cost.
- Fine-tuning comes with the Team plan, alongside endpoints and datasets.
- Serving the adapter is billed as the deployment it runs on.
GPU rates
Related products