Weigths directly in silicon is a bad idea with the way the space is pacing. Just look at chatjimmy.ai it is fast, but on the once good llama3.1-8b but now pretty useless.