Google Cloud has moved its Gemini distillation service into preview. The service takes a large model's answers and reasoning and uses them to train a smaller model, with the goal of cutting latency and per-call costs. During this early-access phase, neither the teacher nor the student model can be swapped out: the teacher is locked to gemini-3.1-pro, and the student to gemini-2.5-flash.
The barriers to entry come first. Project IDs must be added to an allowlist, and the documentation tells users to contact their Google sales representative to apply. Distillation jobs must run in the us-central1 region — there is no second option.
The Dataset Requirements Are Spelled Out in Detail
Training data is a JSONL prompt set stored in Cloud Storage. In each record, the contents field is required and holds the user-side input; systemInstruction is optional and carries the system prompt. Multi-turn conversations alternate between user and model turns.
On scale, the documentation recommends at least 1,000 examples, with a hard cap of 50,000; the source JSONL file cannot exceed 1 GB; each record's input is capped at 8,000 tokens, and the teacher model's output is capped at 24,000 tokens. The service accepts text only — multimodal input is not supported.
On pricing and quotas, the documentation says nothing at all. The service currently sits under pre-GA terms, and Google itself notes that it is not recommended for production use.
Four Steps Folded Into One Job
Distillation itself is an old technique. Doing it yourself meant building the whole pipeline: prompting the teacher model in bulk to generate answers, cleaning and deduplicating that data, fine-tuning the student model on it, and then running a round of evaluation. Every one of those four steps required writing your own code and managing your own compute. The managed service folds them into a single job — users just submit a prompt set.
What you save is engineering effort; what you give up is choice. Both the teacher and the student are locked to specific models, so users can't pair a pricier teacher with the pipeline to push quality further, and they can't feed the output into a custom-architecture small model. The us-central1 restriction, meanwhile, pins down exactly where the data resides — for users with compliance requirements, that constraint matters more than the feature itself.
Same Word, Two Different Treatments
The word “distillation” has taken on a loaded meaning in the US-China context over the past couple of years. Three US agencies have this year named six Chinese companies for distilling US models, and Anthropic has noted in its own reports the large-scale API usage by companies including Alibaba. In those contexts, distillation gets treated as behavior that needs to be held accountable.
The difference lies in the terms of service, not the technical act itself. In Google's service, both the teacher and the student are Google's own models; users run their own prompt sets through it, and the resulting student model serves the same account — the whole process stays within the scope of what's authorized. Cross-vendor distillation — training your own model on another provider's API outputs — is typically banned outright by terms of service. The same sequence of operations is a product feature inside the authorized boundary, and a breach of contract or grounds for litigation outside it.
For developers in China, this service is close to unusable in practice. Getting onto the allowlist requires clearing a sales conversation, and the region is locked to us-central1 — the two restrictions stacked together shut out most China-based projects. What is useful, though, is the set of parameters Google has published: a floor of 1,000 examples, a ceiling of 50,000, and 8,000-token inputs. That is Google's own engineering benchmark for how much data a single distillation run needs, and teams doing similar work elsewhere can use it as a yardstick.
Gaps During the Preview Period
GA timing, pricing, and quota ceilings are all still missing. The documentation doesn't say whether the roster of teacher and student models will expand either. Until that information fills in, the service looks less like a tool to slot into a roadmap and more like an early-access channel for large customers already running Gemini Enterprise.
Sources: Google Cloud official documentation, CocoLoop; teacher and student model versions, region restrictions, dataset fields, and all limits were checked line by line against the original documentation.