Editorial read
In depth · editorial + model
Compressing a large teacher model's behavior into a smaller student to slash serving cost and latency with limited quality loss. It is the primary lever converting expensive frontier models into economically servable products at scale. 2025's low-cost reasoning models showed that distilling large-model behavior into small ones can approach frontier quality at a fraction of the serving cost, forcing a repricing of the field. Over-distillation loses edge-case capability and safety behavior, shipping cheap models that fail silently on hard inputs.