Have you guys seen this? seems like they found a better method to distill models
https://research.nvidia.com/labs/toronto-ai/AlignYourFlow/