Models API — list, describe and download models

List installed models

GET /v1/models returns the models that are installed and allowed by the operator's model lifecycle policy, in OpenAI's list shape. Ids are canonical runtime/family/variant names.

{ "object": "list", "data": [
  { "id": "enginea/llama3.2/3b", "object": "model", "created": 1759650000, "owned_by": "enginea" },
  { "id": "enginea/nomic-embed-text/latest", "object": "model", "created": 1759650000, "owned_by": "enginea" }
] }

GET /v1/models/{id} returns one model, accepting either the canonical id or the wire id (llama3.2:3b); an unknown id answers 404 model_not_found.

Browse the catalogue

GET /v1/models/catalog lists every model the server knows how to get, installed or not, with what you need to choose one:

GET /v1/models/installed returns just the installed keys. Operators can refresh the catalogue sources with POST /v1/models/catalog/refresh.

Download a model

POST /v1/models/pull downloads a model and loads it, streaming progress as Server-Sent Events.

curl -N http://<server>/v1/models/pull -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" -d '{"model_key":"enginea/qwen2.5/7b"}'

event: started
data: {"model_key":"enginea/qwen2.5/7b"}

event: progress
data: {"phase":"downloading","completed":1048576000,"total":4683073536,"bps":52428800,"eta_seconds":69}

event: done
data: {"model_key":"enginea/qwen2.5/7b","duration_ms":91234}

phase is a free-text label from the engine (it varies by engine and step); use completed and total for progress bars. A failure ends with event: error. model_key must be the canonical form. Downloads of the same model by several clients are shared, not repeated.

• The key needs the pull endpoint permission if it has an endpoint allowlist.
• Downloads are allowed in the default federated governance mode. In curated mode only operators install models, and the endpoint answers 403 governance_denied; in open mode clients download models on their own devices.
• Downloading also makes the model the server's active model. On a busy shared server, install models in the app's Models page instead.

Removing models is an operator task in the app's Models page; the API's delete route refuses client requests.

Provenance

GET /v1/server/models/{id}/provenance returns the model card an operator registered — publisher, source, licence, intended use and known limitations — or 404 when none is registered. Review it before approving a model for your organisation.

Questions