Live Models

A model is live when a caller can complete POST /v1/chat/completions against it. That is not Azure-only.

In production (NODE_ENV=production or Cloud Run), serverless is not advertised as live when none of Vertex ADC/project, Together key, or Together BYOK exist. The gateway returns 503 with wholesale_not_configured. Legacy Together-only ids still 503 if Together is missing even when Vertex is up — they are not aliased to Gemma or Gemini.

Public /api/public/pricing marks Together-only rows unavailable without a Together key. Vertex MaaS rows are live when ADC/project is set.

Serverless (Vertex Model Garden)

Wholesale path for this project. Verified HTTP 200 on project-800192c2-3ecc-4889-8f7 via the OpenAI-compatible Vertex endpoint or generateContent (no dedicated GPU):

Images and video (Vertex)

Probed 2026-08-17 on project-800192c2-3ecc-4889-8f7 with ADC. Only 200 ids are advertised.

POST /v1/images/generations and POST /v1/videos/generations. Test in Dashboard → Media.

Llama 3.1 / 3.3 / 4 MaaS (meta/llama-*-maas) still returned 404 after EULA accept. Those ids stay on Together / Groq / Ollama — they are not answered by Gemma or Gemini. Llama 3.1 8B/70B have no MaaS card (self-deploy only). To unlock 3.3 / 4, click Enable on the Model Garden card (not just accept the license):

• Llama 3.3 70B MaaS
• Llama 4 Scout MaaS
• Llama 4 Maverick MaaS

qwen/qwen3-235b-a22b-instruct-2507-maas exists but returned 429 (throttled) — not advertised.

Together overflow (optional)

Local (Ollama)

Prefix any pulled tag with ollama: if you want to force the local adapter.

Vendor APIs (when keyed)

Quick test

bashCopy
curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-oss-20b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

GET /v1/models is the source of truth for what this gateway instance will list.