Large models are powerful, but expensive to run, slow to iterate, and hard to ship. Distillation lets you transfer intelligence from a hosted Azure GPT-5.1 model into a lightweight open-source student that you fully control. In this talk, I’ll show how to generate structured teaching data, design effective prompts and evaluation loops, and fine-tune a smaller model that preserves quality while cutting cost and latency. Walk away with a practical recipe for bringing enterprise-grade reasoning into models you can deploy anywhere.