So which models are you guys using for normal conversations? Llama3 8B instruct seems to be limited in its reasoning. I don’t want to pay a fortune by inference