16 open-model MaaS endpoints retired on Gemini Enterprise Agent Platform: what you need to do now

From 21 October 2026, sixteen open-model endpoints served through Model as a Service (MaaS) on Google Cloud’s Gemini Enterprise Agent Platform are permanently retired. If your code references any of these model IDs, your API calls are failing right now.

This is not a soft deprecation with a grace period. The retirement date is the hard stop: the endpoint is deactivated, the model ID is dead, and there is no fallback.

Which models are gone

The full list of retired endpoints, all deprecated on 21 July 2026 and retired on 21 October 2026, is:

deepseek-ocr-maas
deepseek-r1-0528-maas
deepseek-v3.1-maas
deepseek-v3.2-maas
glm-4.7-maas
glm-5-maas
gpt-oss-20b-maas
kimi-k2-thinking-maas
llama-3.3-70b-instruct-maas
minimax-m2-maas
multilingual-e5-large-instruct-maas
multilingual-e5-small-maas
qwen3-235b-a22b-instruct-2507-maas
qwen3-coder-480b-a35b-instruct-maas
qwen3-next-80b-a3b-instruct-maas
qwen3-next-80b-a3b-thinking-maas

Providers affected span DeepSeek (four endpoints), Alibaba’s Qwen3 family (four), Z.ai’s GLM (two), multilingual E5 embeddings (two), plus single endpoints each from Meta, MiniMax, Moonshot AI’s Kimi, and an OpenAI OSS model.

What “retired” actually means

Google’s lifecycle has two distinct stages, and the distinction matters.

Deprecation (which happened on 21 July 2026) is the warning shot. The endpoint stays live, no new features land, and new usage may be restricted, but existing workloads keep running. It is the platform telling you the clock has started.

Retirement is the termination. From the official deprecations page: “When a model reaches its retirement date, the model endpoint is permanently deactivated and is no longer accessible or supported. API requests calling a retired model ID will fail.”

There was a three-month window to migrate. That window closed on 21 October 2026.

Are you actually affected?

It depends on how you are calling the API. The MaaS retirement applies to the Gemini Enterprise Agent Platform (previously Vertex AI). If your client is configured with a Google Cloud project and location, uses vertexai: true, or runs with the GOOGLE_GENAI_USE_ENTERPRISE=True environment variable, you are on this platform and these retirements apply to you. If you are calling via an AI Studio API key, these particular retirement events do not affect you.

Check your configuration before assuming you are safe.

Your only path forward: self-deploy via Model Garden

Google’s deprecations page lists a recommended self-deploy alternative for each retired model via Model Garden. That is now the only way to continue using these weights on Google Cloud infrastructure.

The tradeoff compared to MaaS is significant.

MaaS was serverless. You sent a request, Google handled the compute, you paid per token, and you never thought about GPUs. Self-deployed models are different in every one of those dimensions. You provision the deployment yourself inside your Google Cloud project and VPC. You own the infrastructure. And critically, self-deployed Model Garden endpoints do not scale to zero. Once a model is deployed to an endpoint, charges accrue continuously whether or not you are sending traffic, until you explicitly undeploy it.

For high-volume, predictable production workloads, self-deployment can produce a lower total cost of ownership over time through optimised per-token pricing at scale. For sporadic or experimental use, the always-on cost model is something to plan for carefully before you deploy.

Are there managed replacements still available?

For some of the retired models, newer MaaS-served successors are available. Two worth noting:

GLM 5.2 from Z.ai became available as a fully managed MaaS endpoint in August 2026. It targets long-horizon agentic and coding tasks and supports a 1 million token context window, making it a direct successor to both glm-4.7-maas and glm-5-maas.

DeepSeek-V4.1-Flash became available as a MaaS endpoint on 14 October 2026. It accepts text and image input, targets agentic and coding tasks, and supports a context window of up to 1 million tokens. It is the obvious managed replacement for the four DeepSeek endpoints that retired this week.

Check the active MaaS catalogue for the full current list.

A note on timing

This retirement arrives one day after a separate shutdown: on 20 October 2026, gemini-2.5-pro, gemini-2.5-flash, and gemini-2.5-flash-lite were also retired on Vertex AI. If you are auditing your integrations this week, check both sets of retirement notices simultaneously rather than treating them as separate exercises.

What to do right now

Search your codebase for any of the 16 model ID strings listed above.
If you find them, decide whether a newer managed MaaS endpoint (such as DeepSeek-V4.1-Flash or GLM 5.2) covers your use case. That is the lowest-friction path.
If no MaaS replacement fits, plan a self-deploy via Model Garden. Factor in the always-on infrastructure cost from day one.
Review the full deprecations page to understand which replacement is recommended for each specific model you use.

The underlying model weights have not disappeared. DeepSeek R1 0528, Qwen3-Coder-480B, MiniMax M2, and the rest can still be run. Google has just stopped running them for you.