Skip to content
agentscamp
Term · Term

Distillation

Distillation trains a smaller model to imitate a larger one — using its outputs as training data to get most of the capability at a fraction of the cost.

Updated Jun 11, 2026
distillationtrainingsmall-modelsmlops

Distillation is training a smaller "student" model on a larger "teacher" model's outputs, transferring most of the teacher's capability on a task into a model that's far cheaper and faster to run.

The pattern in practice: run the frontier model over thousands of representative inputs, capture its outputs (often with reasoning included), curate the best, and fine-tune a small model on the result — synthetic data generation and training in one loop. For narrow tasks (classify, extract, route, rewrite), a distilled small model frequently reaches within a few points of the teacher at a tiny fraction of the per-call cost, which is the whole economics of "GPT-quality at Haiku prices" on your specific workload.

Its boundaries: breadth doesn't transfer (the student learns your task, not general intelligence), quality ceilings inherit from the teacher, and provider terms of service often restrict training on outputs — read them. Where distillation sits against prompting, RAG, and ordinary fine-tuning is mapped in the 2026 decision tree.

Frequently asked questions

How is distillation different from fine-tuning?
Distillation is a kind of fine-tuning where the training data comes from a teacher model: you collect the big model's outputs on your task (a form of synthetic data) and train a small model to reproduce them. Classic fine-tuning uses human-curated examples; distillation manufactures them from a model you wish you could afford to run everywhere.
When does distillation make sense for an LLM product?
When a frontier model nails your task but its cost or latency doesn't fit production scale — and the task is narrow enough for a small model to learn. Classification, extraction, routing, and templated generation distill beautifully; open-ended reasoning distills poorly. Check the model provider's terms first: many restrict using outputs to train competing models.

Related