Google retires 16 open-model endpoints on Vertex AI today
DeepSeek, GLM, Qwen3, Llama, Kimi and MiniMax model-as-a-service endpoints stop working on Vertex AI from 21 October, three months after Google flagged the deprecation.
Google retires 16 open-model endpoints on Vertex AI’s model-as-a-service (MaaS) offering today, 21 October. The list covers four DeepSeek endpoints (V3.1, V3.2, R1-0528 and DeepSeek-OCR), GLM-4.7 and GLM-5, gpt-oss-20b, Kimi K2 Thinking, Llama 3.3 70B, MiniMax M2, two multilingual e5 embedding models, and four Qwen3 variants.
Google’s release notes say the endpoints were deprecated on 21 July, giving three months’ notice, and that after today “the model endpoint is permanently deactivated and is no longer accessible or supported.” Requests calling a retired model ID will simply fail, with no redirect to a replacement. Google points teams that still need these models toward self-deploying them through Model Garden instead of calling the managed endpoint.
Why it matters: if any of these model IDs, deepseek-r1-0528-maas, qwen3-coder-480b-a35b-instruct-maas and so on, are hardcoded into a Vertex AI integration, calls to them stop working from today with no grace period. Anyone running third-party open models through Vertex AI MaaS should check their config now rather than find out from a production error.
The caveat: this only covers Vertex AI’s managed MaaS endpoints. It doesn’t affect these models where they’re self-hosted, run through Model Garden deployments, or served by other providers.