Choose your models

The client never pulls an arbitrary model. Ballast's catalog tells it which models it may pull and the Ollama tag for each, filtered by the VRAM your GPU reported. It never runs anything a renter sends except prompt text.

See and change the list

ballast models
ballast models add qwen3-14b
ballast models remove qwen3-14b

ballast models prints each catalog model with its Ollama tag and one word: serving if you chose it, fits if your GPU could serve it, or needs 16 GB if it could not. A model that is ready shows loaded beside it. A running client picks up a change within about 15 seconds. With no choice made, it serves the largest model that fits.

What happens when a model loads

1. The client pulls the model through Ollama.
2. It runs a short test completion on your machine.
3. Only if that works does it report the model as loaded. Ballast then runs a spot check on it before it is trusted with real jobs.

The first pull of a 20 GB model takes a while. ballast status shows what is still loading.

What each device accepts

On the device's page in your dashboard you choose Serverless only, Dedicated only or Both. Both is the default. A device under 8 GB cannot accept dedicated work. The client does not decide this, the dashboard does. When a GPU is rented, the client keeps that rental's model loaded until the rental ends. More on that in Pause, resume and what you accept.