An 8B model reportedly outperforms a 402B model at tool selection on ToolBench

Researchers at Meta and Carnegie Mellon say their trained Qwen3-8B model outperformed the prompted 402B Llama-4-Maverick at tool selection on ToolBench. DAIR.AI says the 8B model’s ToolBench G3 success rose from 36.2% after supervised training to 41.5% after reinforcement learning. The researchers say the selector also works with different retrievers and new-domain tool catalogs without retraining.