Gemini 3.x Fine-Tuning Still Missing on Vertex AI as 2.5 Pro Deprecation Looms
With gemini-2.5-pro shutting down June 17, 2026, developers can't migrate fine-tuned workloads to Gemini 3.x because supervised fine-tuning remains unavailable on Vertex AI.
With gemini-2.5-pro scheduled to shut down on June 17, 2026, developers are running into a wall: the Gemini 3.x models that are supposed to replace it don't yet support supervised fine-tuning on Vertex AI. The Gemini 3.1 family remains in Preview on Vertex AI, and the feature developers need most — domain-specific training with their own data — is nowhere to be found. The clock is real, and it's ticking loud.
The deprecation notice has been public for weeks, and the migration path seemed straightforward on the surface: move your fine-tuned 2.5 Pro workloads to 3.x, re-train your adapters, update your API calls. Except you can't re-train adapters on 3.x yet because the supervised fine-tuning endpoint simply doesn't exist for that model family on Vertex AI. Developers posting in the Google Developer Forum are explicit: they cannot plan their migration path without this feature. One developer noted that fine-tuning is essential for production workloads in regulated industries — you can't ship a generic model when your use case requires specialized behavior around domain terminology, formatting rules, or compliance constraints.
The practical implications depend on your situation. If you're running Gemini 2.5 Pro in a stateless way — just calling the API with prompts, no fine-tuned adapters — the migration is straightforward: update your model name, test your prompts against 3.x, move on. If you're running fine-tuned models in production, you have roughly ten weeks and no supported path to migrate that fine-tuning to 3.x. Your options in that window are limited: either stick with 2.5 Pro until the feature ships (risky, given the deprecation date), find an alternative provider that supports fine-tuning for comparable models, or invest in prompt engineering to compensate for the lack of specialized training.
This is a pattern that keeps appearing in the Gemini ecosystem: features ship in Preview, get deprecated, and get replaced by something that isn't quite ready before the old thing goes away. The Vertex AI team has been moving fast on the 3.x API surface, but fine-tuning is a different kind of feature — it requires infrastructure for training jobs, evaluation pipelines, and model versioning that can't be rushed. The question isn't whether Google will ship it; it's whether they ship it before June 17. If you're in this situation, it probably makes sense to open a support ticket and get on record with your timeline dependency.