Mycelis
Mycelis
Training & fine-tuning
Train an adapter (LoRA) on your own examples, then deploy the result as a model in your workspace. Mycelis chooses a GPU with enough memory and bills the compute time — nothing more.
No base fee · No minimum term · Pause any time
From data to model
What you get
No training cluster to set up. Pick a base model, supply the data and start the run.
Train a compact adapter on top of an open-source base model instead of retraining the whole model.
Training runs get a GPU with enough memory for the job — you do not size it yourself.
Deploy the finished adapter on its base model and call it through the same endpoint.
You are billed for the time a run actually takes.
Give the tuned model a system prompt, tools and knowledge bases.
Use fine-tuning for style and behaviour, knowledge bases for facts.
How it works
01
Pick an open-source base model and provide your training examples.
02
Mycelis assigns a suitable GPU and trains the adapter.
03
Deploy the adapter and call it like any other model.
Business model
Training is billed by compute time. The deployed model is billed like every hosted model.
Training
Billed by the compute time the run actually uses, at the GPU rate.
Hosting the result
Billed by the minute while the model runs; stop it when nobody needs it.
No package, no minimum term
Nothing to pay between runs.
Common questions
A technique that trains a small adapter instead of the whole model. It is faster, needs less memory and the result is easy to deploy.
No. You choose the base model and provide examples; Mycelis handles the GPU and the run.
Use fine-tuning to teach style, format or behaviour, and a knowledge base when the model needs to look up current facts. Many teams use both.
The compute time of the training run, plus the usual hosting cost while the resulting model runs.
Ready to get started?
Create a free account and start your first fine-tuning run.