Skip to content
Docs

Model Registry

The registry is the list of models this deployment knows about. Registering does not download anything and does not start anything — it records where the weights are, or will come from.

Three ways to register

Section titled “Three ways to register”

Models → Add Model.

  • Repository id — e.g. Qwen/Qwen3-8B. Nothing is fetched until you download it, so you can register a catalogue up front.
  • Local path — a directory of weights or a single .gguf file already on the host. The standard air-gapped workflow, and the way to use an internally mirrored model store.
  • Recommended — a curated list, filtered to what this machine can serve, ordered by what runs best on it. A model absent from that list is not unavailable; it is not servable here, and you can still register it by id or path.

Set an alias

Section titled “Set an alias”

The alias is what GET /v1/models advertises and what applications send. Two reasons it matters:

  • It keeps the raw filesystem path out of the API. Without one, the listing shows the path the server was launched with.
  • It decouples the application from the checkpoint. Serving a newer model under the same alias needs no client change.

Set it when registering, or later with the Models page’s edit action.

What a row carries

Section titled “What a row carries”

Beyond source and alias: detected format and architecture, size, a capability record once probed, and — for models Lifeboat recognizes — the paired speculative-decoding draft model it fetched for you.

Deleting

Section titled “Deleting”

Two different actions, and the distinction is deliberate:

  • Remove from registry — forgets the model. The weights stay on disk. This is the default, and it is reversible by re-registering the path.
  • Delete files from disk — removes the weights too.

Disk deletion is restricted to the approved models directories, and the registry row is removed only after the disk operation succeeds, so a partial failure leaves something to retry rather than an orphaned row pointing at half a model.

For a single .gguf file the parent directory is cleaned up if it becomes empty — so deleting one quantization out of a multi-quant repository leaves its siblings alone.

Replicas on other nodes are not swept. Delete those on each node.

Sources and references

Section titled “Sources and references”