How about smaller more performant models? There’s so much redundancy in language that it should be possible.