Model Card for LoopTool-32B

Model Details

Model Description

The LoopTool-32B model is derived from iterative fine-tuning of Qwen3-32B, with a particular emphasis on enhancing the model’s capabilities in tool invocation.

• Developed by: SJTU, Xiaohongshu
• Model type: Causal Language Models
• Finetuned from model Qwen3-32B: https://huggingface.co/Qwen/Qwen3-32B

Model Sources

• Repository: https://github.com/Rednote-ExperienceAI-Lab/LoopTool
• Paper: https://arxiv.org/abs/2511.09148
• Dataset(Partial) https://huggingface.co/datasets/zhuiguang-ning/LoopTool-23k

Model Performance

The Main Result in BFCL-v3

	Overall	Non-Live	Live	Multi-Turn
Qwen3-8B	66.34	88.81	78.54	33.00
LoopTool-8B	74.93	89.52	84.72	50.88
Qwen3-32B	69.25	88.90	77.83	43.12
LoopTool-32B	79.32	91.83	88.58	57.75

The Main Result in ACEBench (English)

	Overall	Normal	Special	Agent
Qwen3-8B	67.1	70.9	78.0	34.2
LoopTool-8B	73.4	78.0	80.7	43.3
Qwen3-32B	72.2	77.3	76.0	46.7
Kimi-K2-0711	77.4	78.9	81.3	65.0
LoopTool-32B (OpenSource-1st)	77.5	80.5	78.7	64.1

Citation

If you find our work helpful, feel free to give us a cite.

@misc{zhang2025looptool,
      title={LoopTool: Closing the Data-Training Loop for Robust LLM Tool Calls}, 
      author={Kangning Zhang and Wenxiang Jiao and Kounianhua Du and Yuan Lu and Weiwen Liu and Weinan Zhang and Lei Zhang and Yong Yu},
      year={2025},
      eprint={2511.09148},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.09148}, 
}