Welcome to The AI Distillery

The most capable AI in the world is, for most people, something you rent. It lives in a datacenter, answers through an API, and bills you by the token. You don't run it; you query it.

We think that's about to change — and that the mechanism of the change is distillation.

The shift already underway

Distillation is the craft of teaching a small model to think like a large one. For most of its history it was a quiet compression trick. Then, in the last two years, two things happened:

1. Open teacher models got genuinely excellent, so anyone could distill from a strong source without a closed API.
2. Reasoning became distillable. When DeepSeek-R1 generated ~800,000 reasoning traces and used them to train small open students, a 32B model started beating OpenAI's o1-mini on reasoning benchmarks. Frontier capability — not just frontier answers — became something you could run on your own hardware.

That second point is the whole game. It means the future of capable AI doesn't have to be a subscription. It can be a download.

What this site is

The AI Distillery is, first, a knowledgebase and blog — the clearest, most current explanation we can write of how model distillation actually works, from first principles to the on-policy frontier. The field is moving fast and almost nobody is documenting it well. We intend to be the place you come to learn the craft.

We're starting with:

• A primer that assumes no background.
• The conceptual map of the three kinds of knowledge transfer.
• How distillation differs from quantization and pruning — and how they stack.
• A practitioner's toolkit and a guide to running distilled models locally.
• A glossary for every term.

Where it's going

A knowledgebase is the start, not the destination. As the patterns stabilize, we want to build and open-source the tooling that makes distilling a model approachable — and, eventually, a home where the community can share, discover, and trust the small models it distills. Maybe, one day, a marketplace.

There's something almost mythic about all this. We take a vast, inscrutable intelligence and condense it — like drawing a single living thing from a template — into something one person can hold and run and understand. We find that thrilling. But its purpose is deeply human: to put capability back in individual hands, to keep your data yours, to make the future of AI something you participate in rather than merely subscribe to.

Distillation is in its infancy. We're going to document it as it's invented. Come distill with us.

— The AI Distillery