How Distillation works.
Knowledge distillation trains a smaller student model to imitate useful behavior from a larger teacher model. The teacher generates targets, scores, or intermediate signals, and the student learns to reproduce them with less compute.
For builders, distillation can reduce serving cost or enable deployment on smaller hardware, but the result must be evaluated on the real task. A student may preserve common behavior while losing rare skills, calibration, or safety properties. A sound workflow defines the evaluation set first, creates legally and contractually permitted training data, trains the student, and compares quality, latency, and cost before replacing the teacher.
Where it helps.
- 01Reducing inference costs in production
- 02Creating task-specific compact models
- 03Mobile model optimization
- 04Building cheaper alternatives to large models