local llm doesn't need to match SOTA performance in order to be useful.