Choosing exactly what your GPU serves

Your card, your call. You choose which models this machine serves and at which quantization, and the network follows — no ticket, no waiting for us.

Two decisions, kept separate

Which models you serve, and how well you serve them. They are different questions, so they are different settings: you can pin a quality without narrowing your catalogue, or narrow your catalogue without touching quality.

$ mahout config set models_allowed=qwen3.8-27b
$ mahout config set preferred_builds=qwen3.8-27b@ud_iq4_xs

Or the same two choices in your dashboard, on the machine's own row: a checkbox per model, a dropdown per quantization. Only builds your card can actually hold are offered, so you cannot pick something that will not run.

Nine ways to run one model

Qwen3.8 27B ships in nine builds, from 18.4 GiB down to 6.8 GiB. A bigger build answers better; a smaller one answers faster. On a 24 GB card you can run the reference weights, or deliberately run IQ4_XS and serve more requests an hour. Both are legitimate businesses.

Changed while it works

The agent re-reads your settings every minute. Change a quantization from your phone and the machine unloads the old weights and loads the new ones on its own — no restart, no SSH, nothing to remember. Whatever you set is what it goes back to after a reboot.

A pin is a promise we keep

If you pin a build your card cannot hold, we do not quietly serve a smaller one instead. The model is dropped and the agent says why. You asked for a quality, and a silent downgrade would show up weeks later as a benchmark you cannot explain.

What you serve is also what you are paid for — see what you pay, and what a host earns.