LongCat-Flash-Thinking-2601 技术报告
人工智能
2026-02-03 v2
摘要
我们介绍 LongCat-Flash-Thinking-2601，这是一个拥有卓越智能体推理能力的 5600 亿参数开源混合专家 (Mixture-of-Experts, MoE) 推理模型。LongCat-Flash-Thinking-2601 在包括智能体搜索、智能体工具使用和工具集成推理等多个智能体基准测试中，实现了开源模型的最高性能。除了基准测试表现外，该模型在复杂工具交互方面表现出强大的泛化能力，并在噪声较大的真实世界环境中表现出稳健的行为。其高级能力源于一种统一的训练框架，该框架结合了领域并行专家训练与后续融合，并通过从预训练到后训练贯穿数据构建、环境、算法和基础设施的端到端协同设计实现。特别地，模型在复杂工具使用场景中的强大泛化能力，源于我们对环境规模和原则任务构建的深入探索。为优化长尾分布、不平衡的生成以及多轮智能体交互，并实现跨越 10 多千个环境（涵盖 20 多个领域）的稳定训练，我们系统性地扩展了异步强化学习框架 DORA，以实现大规模多环境训练的稳定性和效率。此外，认识到真实任务本质上具有噪声，我们进行了一项系统性分析和噪声模式分解，并设计了针对性的训练程序，显式地将此类不完美纳入训练过程，从而提升了真实世界应用的鲁棒性。为进一步提升复杂推理任务的性能，我们引入了“深度思考模式”，通过大规模并行思考，有效地在推理深度和宽度上进行扩张，从而实现测试时规模化。
引用
@article{arxiv.2601.16725,
title = {LongCat-Flash-Thinking-2601 Technical Report},
author = {Meituan LongCat Team and Anchun Gui and Bei Li and Bingyang Tao and Bole Zhou and Borun Chen and Chao Zhang and Chao Zhang and Chen Gao and Chen Zhang and Chengcheng Han and Chenhui Yang and Chuyu Zhang and Cong Chen and Cunguang Wang and Daoru Pan and Defei Bu and Dengchang Zhao and Di Xiu and Dishan Liu and Dongyu Ru and Dunwei Tu and Fan Wu and Fengcheng Yuan and Fengcun Li and Gang Xu and Guanyu Wu and Guoyuan Lin and Haibin Wang and Hansi Yang and Hao Yang and Haonan Yan and Haoxiang Ma and Haoxing Wen and Hongyan Hao and Hongyin Tang and Hongyu Zang and Hongzhi Ni and Hui Su and Jiacheng Zhang and Jiahong Zhou and Jiahuan Li and Jiaming Wang and Jian Yang and Jianfei Zhang and Jianhao Xu and Jianing Wang and Jiapeng Zhu and Jiaqi Sun and Jiarong Shi and Jiarui Zhao and Jingang Wang and Jinluan Yang and Jinrui Ding and Jinwei Xiao and Jiyuan He and Juncan Xu and Kefeng Zhang and Keheng Wang and Li Wei and Lianhui Ma and Lin Qiu and Lingbing Kong and Lingchuan Liu and Linsen Guo and Mengshen Zhu and Mengxia Shen and Mingyang Zhu and Peiguang Li and Peng Pei and Peng Zhao and Pengcheng Jia and Pengtao Zhang and Ping Liu and Qi Gu and Qiong Huang and Qiyuan Duan and Quanchi Weng and Rongxiang Weng and Rongzhi Zhang and Rumei Li and Shanglin Lei and Shengnan An and Shijun Dai and Shizhe Wu and Shuaikang Liu and Shuang Zhou and Shuo Wang and Songyuan Zhao and Tao Liang and Tianhao Hu and Tianze Chen and Wei Liu and Wei Shi and Wei Wang and Weifeng Tang and Wenjie Shi and Wenlong Zhu and Wentao Chen and Wentao Shi and Xi Su and Xiandi Ma and Xiangcheng Liu and Xiangyu Xi and Xiangyuan Liu and Xiangzhou Huang and Xiao Liu and Xiaodong Cai and Xiaolong Chen and Xiaowei Shi and Xiaoyu Li and Xin Chen and Xingchen Liu and Xuan Huang and Xuezhi Cao and Xunliang Cai and Yan Chen and Yang Bai and Yang Liu and Yang Yang and Yang Zheng and Yanyu Chen and Yaoming Wang and Yaoming Zhu and Yaorui Shi and Yaqi Huo and Yerui Sun and Yi Zhang and Yi-Kai Zhang and Yifan Lu and Yifan Zhao and Yihao Chen and Yitao Zhai and Yongjing Yin and Yongwei Zhou and Youshao Xiao and Yu Wang and Yu Yang and Yuchen Xie and Yuchen Yu and Yuchuan Dai and Yue Xu and Yueqing Sun and Yufei Zhang and Yuhuai Wei and Yulei Qian and Yunfan Liang and Yunke Zhao and Yuwei Jiang and Yuxin Bian and Yuxin Chen and Yuxin Liu and Zeyang Yu and Zhao Yang and Zhengsheng Huang and Zhengyu Chen and Zhijian Liu and Zhikang Xia and Zhimin Lin and Zhiyuan Yao and Zhuofan Chen and Zhuowen Han and Zijian Zhang and Ziran Li and Ziwen Wang and Ziyuan Zhuang},
journal= {arXiv preprint arXiv:2601.16725},
year = {2026}
}