It would be REALLY cool to see this same technique applied to a much more recent OSS model distillation. For example, Mistral 3 14B would be a great target. How efficient can we get inference there?