A better training regimen and better architecture optimizations have allowed smaller models to push above their weight. The leaderboard has many open 7B and 13B models that are comparable with 72B models: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderb...