What's this about?
A Show HN pitching a distill-and-serve pipeline: capture what an expensive frontier model does for your workload, train a smaller model to imitate it, serve that at half the cost. It's the third "same quality, fraction of the price" launch in a week, after Echo's ensemble router.
What exactly is "distillation"?
Distillation trains a small "student" model to imitate a big "teacher" model's outputs — like an apprentice learning the master's finished dishes without the master's years of training. The student stays cheaper because it only needs to mimic the teacher on your kinds of tasks, not everything.
Why does it matter?
The wave of routers, ensembles, and distillers all monetize the same gap: frontier prices are high and workloads rarely need frontier generality. It's also the mechanism labs accuse rivals of using on them — see Friday's Kimi-Claude stylometry story.