Cover Image

LongCat-Flash Technical Report

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Meituan LongCat Team, Bayan, Li, Bei, Lei, Bingye, Wang, Bo, Rong, Bolin, Wang, Chao, Zhang, Chao, Gao, Chen, Zhang, Chen, Sun, Cheng, Han, Chengcheng, Xi, Chenguang, Zhang, Chi, Peng, Chong, Qin, Chuan, Zhang, Chuyu, Chen, Cong, Wang, Congkui, Ma, Dan, Pan, Daoru, Bu, Defei, Zhao, Dengchang, Kong, Deyang, Liu, Dishan, Huo, Feiye, Li, Fengcun, Zhang, Fubao, Dong, Gan, Liu, Gang, Xu, Gang, Li, Ge, Tan, Guoqiang, Lin, Guoyuan, Jing, Haihang, Fu, Haomin, Yan, Haonan, Wen, Haoxing, Zhao, Haozhe, Liu, Hong, Shi, Hongmei, Hao, Hongyan, Tang, Hongyin, Lv, Huantian, Su, Hui, Li, Jiacheng, Liu, Jiahao, Li, Jiahuan, Yang, Jiajun, Wang, Jiaming, Yang, Jian, Tan, Jianchao, Sun, Jiaqi, Zhang, Jiaqi, Fu, Jiawei, Yang, Jiawei, Hu, Jiaxi, Qin, Jiayu, Wang, Jingang, He, Jiyuan, Kuang, Jun, Mei, Junhui, Liang, Kai, He, Ke, Zhang, Kefeng, Wang, Keheng, He, Keqing, Gao, Liang, Shi, Liang, Ma, Lianhui, Qiu, Lin, Kong, Lingbin, Si, Lingtong, Lyu, Linkun, Guo, Linsen, Yang, Liqi, Yan, Lizhi, Xia, Mai, Gao, Man, Zhang, Manyuan, Zhou, Meng, Shen, Mengxia, Tuo, Mingxiang, Zhu, Mingyang, Li, Peiguang, Pei, Peng, Zhao, Peng, Jia, Pengcheng, Sun, Pingwei, Gu, Qi, Li, Qianyun, Li, Qingyuan, Huang, Qiong, Duan, Qiyuan, Meng, Ran, Weng, Rongxiang, Shao, Ruichen, Li, Rumei, Wu, Shizhe, Liang, Shuai, Wang, Shuo, Dang, Suogui, Fang, Tao, Li, Tao, Chen, Tefeng, Bai, Tianhao, Zhou, Tianhao, Xie, Tingwen, He, Wei, Huang, Wei, Liu, Wei, Shi, Wei, Wang, Wei, Wu, Wei, Zhao, Weikang, Zan, Wen, Shi, Wenjie, Nan, Xi, Su, Xi, Li, Xiang, Mei, Xiang, Ji, Xiangyang, Xi, Xiangyu, Huang, Xiangzhou, Li, Xianpeng, Fu, Xiao, Liu, Xiao, Wei, Xiao, Cai, Xiaodong, Chen, Xiaolong, Liu, Xiaoqing, Li, Xiaotong, Shi, Xiaowei, Li, Xiaoyu, Wang, Xili, Chen, Xin, Hu, Xing, Miao, Xingyu, He, Xinyan, Zhang, Xuemiao, Hao, Xueyuan, Cao, Xuezhi, Cai, Xunliang, Yang, Xurui, Feng, Yan, Bai, Yang, Chen, Yang, Yang, Yang, Huo, Yaqi, Sun, Yerui, Lu, Yifan, Zhang, Yifan, Zang, Yipeng, Zhai, Yitao, Li, Yiyang, Yin, Yongjing, Lv, Yongkang, Zhou, Yongwei, Yang, Yu, Xie, Yuchen, Sun, Yueqing, Zheng, Yuewen, Wei, Yuhuai, Qian, Yulei, Liang, Yunfan, Tai, Yunfang, Zhao, Yunke, Yu, Zeyang, Zhang, Zhao, Yang, Zhaohua, Zhang, Zhenchao, Xia, Zhikang, Zou, Zhiye, Zeng, Zhizhao, Su, Zhongda, Chen, Zhuofan, Zhang, Zijian, Wang, Ziwen, Jiang, Zixu, Zhao, Zizhe, Wang, Zongyu, Su, Zunhai
Format:	Preprint
Published:	2025
Subjects:	Computation and Language Artificial Intelligence Distributed, Parallel, and Cluster Computing Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866916957903650816
author	Meituan LongCat Team Bayan Li, Bei Lei, Bingye Wang, Bo Rong, Bolin Wang, Chao Zhang, Chao Gao, Chen Zhang, Chen Sun, Cheng Han, Chengcheng Xi, Chenguang Zhang, Chi Peng, Chong Qin, Chuan Zhang, Chuyu Chen, Cong Wang, Congkui Ma, Dan Pan, Daoru Bu, Defei Zhao, Dengchang Kong, Deyang Liu, Dishan Huo, Feiye Li, Fengcun Zhang, Fubao Dong, Gan Liu, Gang Xu, Gang Li, Ge Tan, Guoqiang Lin, Guoyuan Jing, Haihang Fu, Haomin Yan, Haonan Wen, Haoxing Zhao, Haozhe Liu, Hong Shi, Hongmei Hao, Hongyan Tang, Hongyin Lv, Huantian Su, Hui Li, Jiacheng Liu, Jiahao Li, Jiahuan Yang, Jiajun Wang, Jiaming Yang, Jian Tan, Jianchao Sun, Jiaqi Zhang, Jiaqi Fu, Jiawei Yang, Jiawei Hu, Jiaxi Qin, Jiayu Wang, Jingang He, Jiyuan Kuang, Jun Mei, Junhui Liang, Kai He, Ke Zhang, Kefeng Wang, Keheng He, Keqing Gao, Liang Shi, Liang Ma, Lianhui Qiu, Lin Kong, Lingbin Si, Lingtong Lyu, Linkun Guo, Linsen Yang, Liqi Yan, Lizhi Xia, Mai Gao, Man Zhang, Manyuan Zhou, Meng Shen, Mengxia Tuo, Mingxiang Zhu, Mingyang Li, Peiguang Pei, Peng Zhao, Peng Jia, Pengcheng Sun, Pingwei Gu, Qi Li, Qianyun Li, Qingyuan Huang, Qiong Duan, Qiyuan Meng, Ran Weng, Rongxiang Shao, Ruichen Li, Rumei Wu, Shizhe Liang, Shuai Wang, Shuo Dang, Suogui Fang, Tao Li, Tao Chen, Tefeng Bai, Tianhao Zhou, Tianhao Xie, Tingwen He, Wei Huang, Wei Liu, Wei Shi, Wei Wang, Wei Wu, Wei Zhao, Weikang Zan, Wen Shi, Wenjie Nan, Xi Su, Xi Li, Xiang Mei, Xiang Ji, Xiangyang Xi, Xiangyu Huang, Xiangzhou Li, Xianpeng Fu, Xiao Liu, Xiao Wei, Xiao Cai, Xiaodong Chen, Xiaolong Liu, Xiaoqing Li, Xiaotong Shi, Xiaowei Li, Xiaoyu Wang, Xili Chen, Xin Hu, Xing Miao, Xingyu He, Xinyan Zhang, Xuemiao Hao, Xueyuan Cao, Xuezhi Cai, Xunliang Yang, Xurui Feng, Yan Bai, Yang Chen, Yang Yang, Yang Huo, Yaqi Sun, Yerui Lu, Yifan Zhang, Yifan Zang, Yipeng Zhai, Yitao Li, Yiyang Yin, Yongjing Lv, Yongkang Zhou, Yongwei Yang, Yu Xie, Yuchen Sun, Yueqing Zheng, Yuewen Wei, Yuhuai Qian, Yulei Liang, Yunfan Tai, Yunfang Zhao, Yunke Yu, Zeyang Zhang, Zhao Yang, Zhaohua Zhang, Zhenchao Xia, Zhikang Zou, Zhiye Zeng, Zhizhao Su, Zhongda Chen, Zhuofan Zhang, Zijian Wang, Ziwen Jiang, Zixu Zhao, Zizhe Wang, Zongyu Su, Zunhai
author_facet	Meituan LongCat Team Bayan Li, Bei Lei, Bingye Wang, Bo Rong, Bolin Wang, Chao Zhang, Chao Gao, Chen Zhang, Chen Sun, Cheng Han, Chengcheng Xi, Chenguang Zhang, Chi Peng, Chong Qin, Chuan Zhang, Chuyu Chen, Cong Wang, Congkui Ma, Dan Pan, Daoru Bu, Defei Zhao, Dengchang Kong, Deyang Liu, Dishan Huo, Feiye Li, Fengcun Zhang, Fubao Dong, Gan Liu, Gang Xu, Gang Li, Ge Tan, Guoqiang Lin, Guoyuan Jing, Haihang Fu, Haomin Yan, Haonan Wen, Haoxing Zhao, Haozhe Liu, Hong Shi, Hongmei Hao, Hongyan Tang, Hongyin Lv, Huantian Su, Hui Li, Jiacheng Liu, Jiahao Li, Jiahuan Yang, Jiajun Wang, Jiaming Yang, Jian Tan, Jianchao Sun, Jiaqi Zhang, Jiaqi Fu, Jiawei Yang, Jiawei Hu, Jiaxi Qin, Jiayu Wang, Jingang He, Jiyuan Kuang, Jun Mei, Junhui Liang, Kai He, Ke Zhang, Kefeng Wang, Keheng He, Keqing Gao, Liang Shi, Liang Ma, Lianhui Qiu, Lin Kong, Lingbin Si, Lingtong Lyu, Linkun Guo, Linsen Yang, Liqi Yan, Lizhi Xia, Mai Gao, Man Zhang, Manyuan Zhou, Meng Shen, Mengxia Tuo, Mingxiang Zhu, Mingyang Li, Peiguang Pei, Peng Zhao, Peng Jia, Pengcheng Sun, Pingwei Gu, Qi Li, Qianyun Li, Qingyuan Huang, Qiong Duan, Qiyuan Meng, Ran Weng, Rongxiang Shao, Ruichen Li, Rumei Wu, Shizhe Liang, Shuai Wang, Shuo Dang, Suogui Fang, Tao Li, Tao Chen, Tefeng Bai, Tianhao Zhou, Tianhao Xie, Tingwen He, Wei Huang, Wei Liu, Wei Shi, Wei Wang, Wei Wu, Wei Zhao, Weikang Zan, Wen Shi, Wenjie Nan, Xi Su, Xi Li, Xiang Mei, Xiang Ji, Xiangyang Xi, Xiangyu Huang, Xiangzhou Li, Xianpeng Fu, Xiao Liu, Xiao Wei, Xiao Cai, Xiaodong Chen, Xiaolong Liu, Xiaoqing Li, Xiaotong Shi, Xiaowei Li, Xiaoyu Wang, Xili Chen, Xin Hu, Xing Miao, Xingyu He, Xinyan Zhang, Xuemiao Hao, Xueyuan Cao, Xuezhi Cai, Xunliang Yang, Xurui Feng, Yan Bai, Yang Chen, Yang Yang, Yang Huo, Yaqi Sun, Yerui Lu, Yifan Zhang, Yifan Zang, Yipeng Zhai, Yitao Li, Yiyang Yin, Yongjing Lv, Yongkang Zhou, Yongwei Yang, Yu Xie, Yuchen Sun, Yueqing Zheng, Yuewen Wei, Yuhuai Qian, Yulei Liang, Yunfan Tai, Yunfang Zhao, Yunke Yu, Zeyang Zhang, Zhao Yang, Zhaohua Zhang, Zhenchao Xia, Zhikang Zou, Zhiye Zeng, Zhizhao Su, Zhongda Chen, Zhuofan Zhang, Zijian Wang, Ziwen Jiang, Zixu Zhao, Zizhe Wang, Zongyu Su, Zunhai
contents	We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat
format	Preprint
id	arxiv_https___arxiv_org_abs_2509_01322
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	LongCat-Flash Technical Report Meituan LongCat Team Bayan Li, Bei Lei, Bingye Wang, Bo Rong, Bolin Wang, Chao Zhang, Chao Gao, Chen Zhang, Chen Sun, Cheng Han, Chengcheng Xi, Chenguang Zhang, Chi Peng, Chong Qin, Chuan Zhang, Chuyu Chen, Cong Wang, Congkui Ma, Dan Pan, Daoru Bu, Defei Zhao, Dengchang Kong, Deyang Liu, Dishan Huo, Feiye Li, Fengcun Zhang, Fubao Dong, Gan Liu, Gang Xu, Gang Li, Ge Tan, Guoqiang Lin, Guoyuan Jing, Haihang Fu, Haomin Yan, Haonan Wen, Haoxing Zhao, Haozhe Liu, Hong Shi, Hongmei Hao, Hongyan Tang, Hongyin Lv, Huantian Su, Hui Li, Jiacheng Liu, Jiahao Li, Jiahuan Yang, Jiajun Wang, Jiaming Yang, Jian Tan, Jianchao Sun, Jiaqi Zhang, Jiaqi Fu, Jiawei Yang, Jiawei Hu, Jiaxi Qin, Jiayu Wang, Jingang He, Jiyuan Kuang, Jun Mei, Junhui Liang, Kai He, Ke Zhang, Kefeng Wang, Keheng He, Keqing Gao, Liang Shi, Liang Ma, Lianhui Qiu, Lin Kong, Lingbin Si, Lingtong Lyu, Linkun Guo, Linsen Yang, Liqi Yan, Lizhi Xia, Mai Gao, Man Zhang, Manyuan Zhou, Meng Shen, Mengxia Tuo, Mingxiang Zhu, Mingyang Li, Peiguang Pei, Peng Zhao, Peng Jia, Pengcheng Sun, Pingwei Gu, Qi Li, Qianyun Li, Qingyuan Huang, Qiong Duan, Qiyuan Meng, Ran Weng, Rongxiang Shao, Ruichen Li, Rumei Wu, Shizhe Liang, Shuai Wang, Shuo Dang, Suogui Fang, Tao Li, Tao Chen, Tefeng Bai, Tianhao Zhou, Tianhao Xie, Tingwen He, Wei Huang, Wei Liu, Wei Shi, Wei Wang, Wei Wu, Wei Zhao, Weikang Zan, Wen Shi, Wenjie Nan, Xi Su, Xi Li, Xiang Mei, Xiang Ji, Xiangyang Xi, Xiangyu Huang, Xiangzhou Li, Xianpeng Fu, Xiao Liu, Xiao Wei, Xiao Cai, Xiaodong Chen, Xiaolong Liu, Xiaoqing Li, Xiaotong Shi, Xiaowei Li, Xiaoyu Wang, Xili Chen, Xin Hu, Xing Miao, Xingyu He, Xinyan Zhang, Xuemiao Hao, Xueyuan Cao, Xuezhi Cai, Xunliang Yang, Xurui Feng, Yan Bai, Yang Chen, Yang Yang, Yang Huo, Yaqi Sun, Yerui Lu, Yifan Zhang, Yifan Zang, Yipeng Zhai, Yitao Li, Yiyang Yin, Yongjing Lv, Yongkang Zhou, Yongwei Yang, Yu Xie, Yuchen Sun, Yueqing Zheng, Yuewen Wei, Yuhuai Qian, Yulei Liang, Yunfan Tai, Yunfang Zhao, Yunke Yu, Zeyang Zhang, Zhao Yang, Zhaohua Zhang, Zhenchao Xia, Zhikang Zou, Zhiye Zeng, Zhizhao Su, Zhongda Chen, Zhuofan Zhang, Zijian Wang, Ziwen Jiang, Zixu Zhao, Zizhe Wang, Zongyu Su, Zunhai Computation and Language Artificial Intelligence Distributed, Parallel, and Cluster Computing Machine Learning We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat
title	LongCat-Flash Technical Report
topic	Computation and Language Artificial Intelligence Distributed, Parallel, and Cluster Computing Machine Learning
url	https://arxiv.org/abs/2509.01322