This is the key aspect for future development of models - small instant reasoning, ideally on device that funnels through tho a larger model for reasoning.