Reference · Glossary
Distillation
Training a **smaller "student" model** to mimic a larger "teacher" model's outputs — keeping much of the quality at lower cost.
When to use
Deploying fast chatbots, on-device assistants, or high-volume APIs where the biggest model is too slow or expensive.
When not to
Cutting-edge research tasks where only frontier models pass evals — distill after you know the teacher works.
Example
A big model generates ideal answers on 50k prompts; a small model learns to match those answers for production traffic.