Model distillation
A technique for compressing a large, capable AI model into a smaller, faster one by training the small model to imitate the big one. Distillation is how much of a model's ability is packed into something cheap enough to run at scale or on a phone.
The most capable AI models are also the most expensive to run, which is a problem if you want to serve millions of users or run a model on a device in someone’s pocket. Model distillation is one answer. A large, powerful model — the “teacher” — is used to train a much smaller “student” model, which learns to reproduce the teacher’s outputs. The student ends up far lighter, yet keeps a surprising amount of the teacher’s ability.
The gain is efficiency. A distilled model needs less compute, responds faster, and costs less to operate, which makes it practical for high-volume services or for running directly on phones and other hardware where a giant model could never fit. Some capability is inevitably lost in the shrinking, but for many everyday tasks the smaller model is more than good enough.
Distillation has become a common step between research and product: build the best model you can, then distil it down to something you can actually afford to deploy. It also features in debates about open models, since a capable open-weight model can be used as a teacher to train others.