D
Member
Distilling a frontier model into a small one is standard practice now, but I rarely see honest numbers on what gets lost: reasoning depth, rare knowledge, calibration. If you have distilled for production, what did you measure, and what surprised you about the trade-offs?