It's really task-dependent, text summarization and grammar corrections are fine with local models. I posit any tasks that are 'arms race-y' (image generation, creative text generation) are going to be offloaded to servers, as there's no 'good enough' bar above which they can't improve.