Training a small model to imitate a big one.
Distillation uses a large "teacher" model to generate outputs that train a much smaller "student", transferring most of the behaviour at a fraction of the serving cost.
The resulting small model is cheaper, faster and often good enough for a narrow task — which is what makes on-device and high-volume AI economically viable.
It is also contentious: distilling a competitor's model is prohibited by most frontier providers' terms.