Skip to main content
Inspect a named model created by a completed fine-tune:
Read production, candidate, previous and versions. Each version summary names its adapter, status, serving state, gate and held-out metrics. Pointers can be null, especially before the first promotion.

​
Choose a reference

A named response reports both the name you requested and the exact version that answered. During evaluation, set fallback: "none" if another model’s result would invalidate the comparison.

​
Promote after review

When the candidate’s gate and your own review are satisfactory:
Promotion verifies, loads and probes the target before moving production. The old production becomes previous. A failed gate blocks ordinary promotion; force: true is an explicit override and does not fix the measured regression. Inspect the evidence before choosing it. A standalone adapter can be loaded with /v1/models/{model_id}/deploy. Deploying an adapter does not itself choose a named model’s production pointer.

​
Roll back

Rollback loads and probes the previous version, then swaps it with production. There must be a previous version. Calling rollback twice swaps again; it is not a repeat-safe recovery command. If a promotion or rollback response is lost, read the model before deciding what to do next.

​
Warmth and memory

keep_warm describes whether production is pinned to stay loaded where residency is managed. state can be loaded, loading, asleep or null when not tracked. See models and fallback for waking a cold version. memory.replay and memory.holdout summarize sampled prior data. New held-out examples are remembered when a job is created; training examples join replay when their version is promoted. This helps compare later versions against older tasks, but it is still a sample and needs your regression checks.