Skip to content
Technology

Distillation

A technique where a smaller 'student' model learns to replicate the behavior of a larger 'teacher' model. Produces compact models that retain most of the teacher's capability while being faster, cheaper, and deployable on smaller devices.

Why it matters

Big models are accurate but slow and expensive to run. Distillation lets companies capture most of that quality in a model small enough to run on a phone, a laptop, or a cheap server. That's why you can now get useful AI responses instantly and often for free. Many of the fast, low-cost models you compare on aggregators started as distilled versions of a heavyweight teacher.

A concrete example

Suppose a company has a giant model that answers customer questions brilliantly but costs too much to run at scale. They train a small student model on the big model's answers to thousands of questions. The student learns to mimic those responses closely. Now it handles routine support tickets at a fraction of the cost, and the expensive teacher is reserved only for the trickiest cases.

How to use it

Distilled models are why a small, fast, cheap model can be surprisingly capable at a narrow task. When choosing, this is the reason to test rather than assume: a distilled small model can outperform a larger general one on the work it was distilled for, while being clearly worse outside it. If you are running your own, distillation is the standard route to a deployable size. The idea goes back to Distilling the Knowledge in a Neural Network.

The common mistake

Expecting the small model to inherit the large one's range. It inherits the behaviour it was trained to copy, so it holds up on the target task and degrades faster than expected on anything adjacent.

Related terms

Put Distillation into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.