Skip to main content

Custom models

Besides the preloaded catalog, an Owner can deploy custom models on a Mother machine: your own fine-tune, a model Mother does not ship yet, or a specialized engine build.

How deployment works​

  1. Prepare the model as a container image with the weights and a manifest describing what it serves (chat, embeddings, …) and its hardware requirements.
  2. Register it on the machine with Mother's BYO-model tooling (mother-byo-model on the node). The manifest tells Mother how to run the model and where it came from.
  3. Use it like any other model — it appears in Models and in Genesis model pickers, and it runs in the same worker runtime as catalog models.

Rules and constraints​

  • Custom models run in the same isolated container runtime as catalog models; they get their own resources and do not reach node services.
  • Removing or updating a custom model is a registry operation; Mother guards against downgrading a model that sessions are actively using.
  • Custom models are per-machine. They do not sync to other machines in a cluster.

Coming soon​

A guided UI for adding custom models is planned. Until then, deployment is a command-line procedure for Owners — ask for the current runbook if you need the exact steps.