Can someone specialized in applied machine learning explain how this is useful? In my opinion, general-purpose models are only useful if they're large, as they are more capable and produce more accurate outputs for certain tasks. For on-device models, fine-tuned ones for specific tasks have greater precision with the same size.