I've been testing small LLM models... Fijik1.5 with 380M Active models is very fast, but it thinks too long and the results are only semantically close. I hope in a few months we'll see great models of these sizes!