There's a huge range of CPU and memory requirements to run these models. The lower end of the range can be run locally but performance is noticeably worse. I'm guessing this will change as more people are looking at efficiency now. Specialized hardware like Apple's Neural Engine on the MacBook Pro may also help.