Gemini Distillation Service (17-minute read)

Gemini Distillation Service (Distillation) is a service that trains a smaller, more efficient 'Student' model using the outputs and reasoning patterns of a larger, more powerful 'Teacher' model. This helps maintain the high performance of frontier models while optimizing latency and cost for specific enterprise use cases.

Unlike conventional SFT (Supervised Fine-Tuning), which uses only the final text output, Distillation utilizes both of the following:

• Teacher responses: The final text output
• Raw thoughts: The internal reasoning path generated by the teacher model

The models currently supported in the early access stage are gemini-3.1-pro as the teacher model and gemini-2.5-flash as the student model.

This service is recommended in the following situations:

• High-volume and low-latency applications: When Pro-level reasoning capability is needed, but strict SLAs or budget constraints require using a Flash-level model
• Lack of ground-truth data: When there is a large amount of user prompt data, but insufficient resources to generate the high-quality ground-truth data needed for standard SFT
• Complex reasoning tasks: When performing tasks that a base Flash model struggles with, such as multi-step logic, technical document summarization, or complex coding

The key point is that a Prompt-only dataset can be used when preparing the dataset. Since the teacher model directly generates the target outputs during the Distillation process, users do not need to provide separate ground-truth answers.