Knowledge distillation

Training a smaller student model to imitate a larger teacher model's outputs so it reaches near-teacher quality at far lower inference cost and latency.

grounded in: Trend theme 'Cheaper/faster model migration: teams re-platforming agents onto newer models for measurable speed and cost wins' + distillation as an established model-compression technique in the infer

Connected concepts

Quantization, Mixture-of-Experts, Speculative decoding, Model migration, Open-weight LLMs