Google Cloud Launches Gemini Distillation Service in Early Access, Letting Enterprises Train Faster, Cheaper AI Models
Summary
Google Cloud launches Gemini Distillation Service in early access, allowing enterprises to train smaller, faster, and cheaper AI models by transferring knowledge from a powerful 'teacher' model to a leaner 'student' model — no manually labeled data required.
Key Points
- Google Cloud's Gemini Distillation Service is now in early access, enabling users to train a smaller, more efficient 'student' model (Gemini 2.5 Flash) by leveraging the outputs and internal reasoning patterns of a larger 'teacher' model (Gemini 3.1 Pro), delivering lower latency and cost for enterprise use cases.
- Unlike standard supervised fine-tuning, the service uses prompt-only datasets in JSONL format — requiring no manually labeled ground-truth answers — with the teacher model automatically generating target outputs during training, supporting a minimum of 1,000 examples and a maximum of 50,000 examples up to 1 GB.
- Distillation jobs must run in the us-central1 region via the Agent Platform REST API, with configurable hyperparameters such as epochCount and learningRateMultiplier, and completed models are automatically registered in the Model Registry with dedicated prediction endpoints for evaluation; access requires project allowlisting with a default concurrent quota of 4.