Well beside for roleplaying, local models are a bit useless, especially the ones running on end user class hardware.

Maybe when chip would infer at 10-20x the speed of a 5090 RTX, otherwise it's really..bad.