Models

Catalog aliases currently served by Mifr.

Mifr owns the catalog. A provider advertises a subset and lazy-loads slots; the coordinator publishes the same JSON as public GET /v1/catalog (on the coordinator, not the gateway) and never decrypts a prompt to consult it.

Alpha aliases

Alias	Hugging Face pin	Min unified memory	License
qwen3-4b (default)	mlx-community/Qwen3-4B-4bit	10 GB	Apache-2.0
qwen3-8b	mlx-community/Qwen3-8B-4bit	14 GB	Apache-2.0

qwen3-4b is the join-test alias by decision, not a paid catalog. Qwen3 is a thinking model, so short max_tokens often returns only <think>….

Two lists that are not the same

Endpoint	Host	Auth	Meaning
GET /v1/models	gateway	Bearer	Catalog aliases currently servable at the gateway trust floor (the union of live, qualified Macs). OpenAI {object:"list",data:[…]}.
GET /v1/catalog	coordinator	public	Full coordinator-owned catalog (aliases, HF ids, revision, SHA-256), independent of who is online.
GET /v1/providers	gateway	Bearer	Live Macs at the floor (id, model_ids, trust, busy, waiting, and serial when the rung is hardware).

A 200 on /v1/models that only lists qwen3-4b means no qualified Mac is advertising qwen3-8b right now, not that the alias has been removed.

Hugging Face ids still resolve on Verified (mlx-community/Qwen3-4B-4bit maps to qwen3-4b). mifr-local maps to the catalog default. Unknown names return 404 model_not_found with mifr_action: pick_model and do not burn a GPU session.

The JSON model field on the completion echoes the request. The alias actually served is x-mifr-model / mifr_receipt.model_id.

Pinning a Mac

The optional body field provider (or provider_key_id) takes an exact provider id, or a unique prefix of at least 8 characters. A pin does not hop or hedge. If that Mac is offline or below the floor, the call returns 503 provider_unavailable. When you omit it, the gateway picks idle-first near-tie.