Vertex AI (EU)

What it is

The gemini/provider reaches Google through the public Gemini API (Google AI Studio) — fast to set up, but Google doesn't commit to serving any particular request from any particular region. vertex/reaches the exact same Gemini models a different way: through Vertex AI, Google Cloud's enterprise inference product, from a GCP project Mizan owns and operates (mizan-vertex-ai). Vertex lets you pin a project to a specific region — Mizan's is pinned to europe-west4 (the Netherlands), so a vertex/ call never leaves the EU.

curl https://api.app-mizan.com/v1/chat/completions \
  -H "Authorization: Bearer mizan_rt_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vertex/gemini-2.5-pro",
    "messages": [{ "role": "user", "content": "Say hi in five words." }]
  }'

No code change beyond the model prefix — same endpoint, same Mizan key, same OpenAI-shaped request and response either way.

Why EU, and why only EU for now

A single fixed region is the whole point: it's a real, verifiable hosting guarantee for teams that need one — not a legal opinion, an actual GCP region setting on the project that serves your call. Vertex supports plenty of other regions too; Mizan started with EU because that's the residency requirement that comes up today. Nothing about the design is EU-specific — a second region is a config change (a new DataResidencyvalue plus a second pinned Vertex project), not new infrastructure — it's just not built until there's a real second requirement to build it for.

How it's billed

Platform-only for now: every vertex/ call bills your Mizan wallet, at the same per-token price as the equivalent gemini/model. There's no BYOK connection type for Vertex yet — connecting your own GCP project so Vertex calls bill your own Google Cloud bill instead is on the roadmap, not shipped. See BYOK for how that model works for every other provider today.

Supported models

Gemini only, for now — the same model ids as the gemini/ provider (e.g. gemini-2.5-pro works as both gemini/gemini-2.5-pro and vertex/gemini-2.5-pro— same model, different hosting guarantee). Vertex also offers other model families as Model Garden options — Anthropic Claude, xAI Grok, DeepSeek, Meta Llama, Mistral, and Qwen are all servable through the same Google Cloud project — but each of those speaks its own native request shape rather than Gemini's OpenAI-compatible one, and none of them are wired up in Mizan yet. Routable model ids always match what Models (or GET /v1/models) shows as active under the vertexprovider — check there rather than assuming a Gemini id you've seen elsewhere is available.

Finding EU-resident models

The Models page has a “Data residency” filter (next to the provider filter) that narrows the catalog to EU-pinned models only — right now, that's the three Vertex Gemini entries above. Every model carries this as a real field (data_residency in GET /v1/models), not just a UI label — "global" for everything else (no residency guarantee either way, not a claim it leaves the EU) and "eu" for the models actually pinned to one.