good point I have the feeling larger models (20b+) rely too much about their stored knowledge and sometimes fail to use tools because they think they know the answer. smaller specialized tool calling models could be the smart route for the future