what about benchmark against smaller or bigger models? 9B looks too small for llm-level decisions.