One big one is speculative decoding for larger models, the more “accurate”* the smaller model, the more the speed up on the bigger model

* as in matches the token that the larger model would output