[Feature]: Support Gemma4-26B-A4B-IT MaaS on Vertex AI

[Nyarn-WTF]

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

The Feature

Please add support for Gemma 4 26B A4B IT (gemma-4-26b-a4b-it-maas), a multimodal model available on Google Vertex AI via Model as a Service (MaaS).

• Model ID: gemma-4-26b-a4b-it-maas
• Provider: Google (via Vertex AI MaaS)
• Launch stage: Experimental (released April 3, 2026)
• Inputs: Text, Documents, Images
• Output: Text
• Context length: 256,000 tokens
• Max output: 128,000 tokens
• Supported capabilities: Function calling, Structured output
• Docs: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/maas/google/gemma-4-26b-a4b-it

Motivation, pitch

I'm building an application on Vertex AI and would like to use Gemma 4 26B A4B IT through LiteLLM's unified interface. This model is already available as a managed service on Vertex AI MaaS (similar to how other open models like Llama and DeepSeek are supported), so adding it would allow LiteLLM users to seamlessly access this model alongside other Vertex AI models.

Since LiteLLM already supports other Vertex AI MaaS models (e.g., vertex_ai/meta/llama-4-scout-17b-16e-instruct-maas), I expect this would follow the same pattern with the model string vertex_ai/google/gemma-4-26b-a4b-it-maas.

What part of LiteLLM is this about?

Proxy

LiteLLM is hiring a founding backend engineer, are you interested in joining us and shipping to all our users?

No

Twitter / LinkedIn details

No response

[owtaylor]

It's possible to use this via the openai/ provider:

OPENAI_API_KEY=$(gcloud auth print-access-token)
OPENAI_API_BASE=https://aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/global/endpoints/openapi
# OPENAI_API_BASE=https://${LOCATION}/aiplatform.googleapis.com/v1/projects/${VERTEXAI_PROJECT}/locations/${LOCATION}/endpoints/openapi
MODEL=openai/google/gemma-4-26b-a4b-it-maas

Fixing this in litellm properly would probably mean convincing this model to go through the "partner_models" path like gpt-oss or llama3. So that would mean changing llms/vertex_ai/vertex_ai_partner_models/main.py:VertexAIPartnerModels.is_vertex_partner_model. (And maybe somehow excluding it from being detected as a gemma model)

Then you'd need to figure out if there is stuff that should be done in an OpenAIGPTConfig subclass like VertexAIGPTOSSTransformation or VertexAILlama3Config. Though it does (as above) work at least basically without that.

cc: @Iana-Kasimova-TR

[Iana-Kasimova-TR]

@owtaylor Thanks, that makes sense. I’d like to take this on and open a PR

I’ll look into routing vertex_ai/google/gemma-4-26b-a4b-it-maas through the existing Vertex partner model path, similar to GPT-OSS / Llama3, starting with VertexAIPartnerModels.is_vertex_partner_model in llms/vertex_ai/vertex_ai_partner_models/main.py.

and I will check whether Gemma MaaS needs its own OpenAIGPTConfig subclass or whether the generic OpenAI-compatible behavior is enough once detection/routing is fixed

and include model metadata and tests to confirm that users can call this model thorough "vertex_ai/google/gemma-4-26b-a4b-it-maas"

[Nyarn-WTF]

Great to see this merged! Thank you to the contributors and maintainers for making this happen. Much appreciated!

[mvrodrig]

Hi @mateo-berri , just wanted to follow up on the timeline for this fix landing in an official release.

PR #28846 was merged on 2026-05-28 into the litellm_oss_staging_28_05_26 branch, but as of today (2026-06-05) the fix doesn't appear to be included in any official release — including the latest v1.87.1. We verified by checking PartnerModelPrefixes in vertex_ai_partner_models/main.py across tags up to v1.87.1.

Thanks!