feat: add google-vertex-openai provider for Vertex AI partner models

[bluet]

Add Vertex AI Partner Models with [provider] Overrides

Adds partner/open models served via Google Vertex AI's OpenAI-compatible endpoint, consolidated under the existing google-vertex provider using per-model [provider] npm overrides (same pattern as Azure's Claude models).

Models Added/Updated

Model	ID	Context	Output	Cost In/Out	Modalities
GLM-4.7 (updated)	zai-org/glm-4.7-maas	200K	128K	$0.60/$2.20	text, pdf
DeepSeek V3.1	deepseek-ai/deepseek-v3.1-maas	163K	32K	$0.60/$1.70	text, pdf
Llama 4 Maverick	meta/llama-4-maverick-17b-128e-instruct-maas	524K	8K	$0.35/$1.15	text, image
Llama 3.3 70B	meta/llama-3.3-70b-instruct-maas	128K	8K	$0.72/$0.72	text
Qwen3 235B	qwen/qwen3-235b-a22b-instruct-2507-maas	262K	16K	$0.22/$0.88	text

Approach

Per maintainer feedback ("Ideally we don't make 3 different providers for vertex"), all models live under providers/google-vertex/models/ in vendor-namespaced subdirectories. Each uses:

[provider]
npm = "@ai-sdk/openai-compatible"

This follows the established pattern (e.g., Azure's claude-haiku-4-5.toml uses [provider] npm = "@ai-sdk/anthropic" to override the provider's default npm package).

Verification

• All specs verified against official Google Cloud documentation:
  • Vertex AI Pricing
  • GLM-4.7 Docs
  • DeepSeek V3.1 Docs
  • Llama 4 Maverick Docs
  • Llama 3.3 Docs
  • Qwen3 Docs
• bun validate passes (exit code 0)

GLM-4.7 Corrections from Original

Fixed specs that differed from official docs:

• context: 204,800 → 200,000
• output: 131,072 → 128,000
• release_date: 2025-12-22 → 2026-01-06
• Added pdf input modality
• Added structured_output = true

Related

• Depends on opencode PR #10303 for runtime support of [provider] overrides with @ai-sdk/openai-compatible

[rekram1-node]

I think im going to consolidate all the google vertex stuff under a single provider and override npm and api fields using [provider] section per model as needed.

This will require additional tweaks in opencode tho

[bluet]

@rekram1-node anything I can do/help with to accelerate this?

I've also created a related PR in opencode anomalyco/opencode#10742

[opencode-agent]

User bluet does not have write permissions

github run

[bluet]

@rekram1-node I've closed my opencode PR #10742 in favor of the existing PR #10303 by @leehack, which takes the same consolidation approach you mentioned (using protocol: "openapi" within the existing google-vertex provider).

For this models.dev PR, what would you like me to do?

Options:

1. Close this PR - if you're planning to handle the consolidation yourself
2. Update this PR - to use the protocol: "openapi" field format that matches PR #10303's approach (move models back under google-vertex with appropriate overrides)
3. Keep as-is temporarily - until the opencode PR #10303 is merged and we know the exact model schema

Let me know what works best for you!

[bluet]

@rekram1-node Restructured per your feedback — all models are now consolidated under the existing google-vertex provider instead of creating a separate google-vertex-openai provider.

Each partner model uses a [provider] override to specify npm = "@ai-sdk/openai-compatible" (same pattern as Azure's Claude models), while staying under providers/google-vertex/models/ in vendor-namespaced subdirectories.

Also corrected all specs against official Google Cloud docs — the original values had some discrepancies (context windows, output limits, pricing).

Ready for review when you have time!

[rekram1-node]

ill look over tmr, sorry for delays been crazy

[rekram1-node]

So I see what you're doing but there is a pretty critical bug here, none of these models have a base url...

You need:

[provider]
npm = ....
api = ... // THIS ONE

And the tricky thing here is we havent actually introduced this pattern yet and its why google vertex anthropic was originally separate provider

What api should look like is also tied to the env vars, a lot of the apis would look something like this?

https://${ENDPOINT}/v1/projects/${PROJECT_ID}/locations/${REGION}/endpoints/openapi

But instead of ${PROJECT_ID} it should be consistent w/ the vars stated in the provider.toml so for example:

GOOGLE_VERTEX_PROJECT

And then in opencode it should be handling the replacing of these values.

[bluet]

Good catch! Added the api field to all partner models:

[provider]
npm = "@ai-sdk/openai-compatible"
api = "https://${GOOGLE_VERTEX_ENDPOINT}/v1/projects/${GOOGLE_VERTEX_PROJECT}/locations/${GOOGLE_VERTEX_LOCATION}/endpoints/openapi"

Note: I used ${GOOGLE_VERTEX_PROJECT} rather than ${GOOGLE_CLOUD_PROJECT} because the merged loadBaseURL() in opencode PR #10303 resolves ${GOOGLE_VERTEX_PROJECT} through the full fallback chain in googleVertexVars() (GOOGLE_VERTEX_PROJECT → options.project → GOOGLE_CLOUD_PROJECT → GCP_PROJECT → GCLOUD_PROJECT), whereas ${GOOGLE_CLOUD_PROJECT} would only resolve via direct Env.get() and miss users who set GOOGLE_VERTEX_PROJECT instead.

bun validate passes. Also congrats on merging opencode PR #10303!