Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2505.15431 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866913925812977664 |
|---|---|
| author | Tencent Hunyuan Team Liu, Ao Zhou, Botong Xu, Can Zhou, Chayse Zhang, ChenChen Xu, Chengcheng Wang, Chenhao Wu, Decheng Wu, Dengpeng Jiao, Dian Du, Dong Wang, Dong Zhang, Feng Lian, Fengzong Xu, Guanghui Zhang, Guanwei Wang, Hai Luo, Haipeng Hu, Han Xu, Huilin Wu, Jiajia Zhu, Jianchen Yan, Jianfeng Zhu, Jiaqi Zhang, Jihong Xue, Jinbao Xia, Jun Zheng, Junqiang Liu, Kai Zhang, Kai Zheng, Kai Li, Kejiao Wang, Keyao Jiang, Lan Liu, Lixin Wu, Lulu Huang, Mengyuan Yu, Peijie Wang, Peiqi Wang, Qian Xiang, Qianbiao Liu, Qibin Sun, Qingfeng Guo, Richard Xie, Ruobing Yang, Saiyong Chen, Shaohua Hu, Shihui Li, Shuai Li, Shuaipeng Chen, Shuang Zheng, Suncong Yang, Tao Zhang, Tian Yu, Tinghao Han, Weidong Liu, Weijie Zhou, Weijin Wang, Weikang Chen, Wesleye Feng, Xiao Ren, Xiaoqin Sun, Xingwu Kuang, Xiong Huang, Xuemeng Cao, Xun Chen, Yanfeng Du, Yang Yang, Zhen Tao, Yangyu Deng, Yaping Shen, Yi Hong, Yigeng Chen, Yiqi Huang, Yiqing Deng, Yuchi Mao, Yue Wang, Yulong Zeng, Yuyuan Xu, Zenan Kang, Zhanhui Zhao, Zhe Yan, ZhenXiang Fang, Zheng Hu, Zhichao Chen, Zhongzhi Li, Zhuoyu Li, Zongwei Yan, Alex Liang, Ande Liu, Baitong Pan, Beiping Xing, Bin Wu, Binghong Qu, Bingxin Ni, Bolin Wu, Boyu Li, Chen Jiang, Cheng Zhang, Cheng Liu, Chengjun Yang, Chengxu Xu, Chengzhong Wang, Chiyu Zha, Chong Yi, Daisy Wang, Di Lu, Fanyang Chen, Fei Liu, Feifei Zheng, Feng Yu, Guanghua Li, Guiyang Wang, Guohua Lin, Haisheng Liu, Han Wang, Han Fei, Hao Lu, Hao Jiang, Haoqing Sun, Haoran Zhu, Haotian Dai, Huangjin Chen, Huankui Feng, Huawen Cai, Huihui Peng, Huxin Lv, Jackson Shi, Jiacheng Bu, Jiahao Li, Jianbo Hu, Jianglu Guan, Jiangtao Xu, Jianing Cai, Jianwei Zhang, Jiarong Song, Jiawei Jiang, Jie Liu, Jie Yang, Jieneng Zhang, Jihong lv, Jin Zhao, Jing Li, Jinjian Liu, Jinxing Zhao, Jun Guo, Juntao Wang, Kai Wu, Kan Fu, Lei He, Lei Wang, Lei Liu, Li Dong, Liang Zhan, Liya Cheng, Long Xu, Long Zheng, Mao Liu, Meng Hu, Mengkang Chen, Nanli Chen, Peirui He, Peng Pan, Pengju Wei, Pengzhi Yang, Qi Yi, Qi Wang, Roberts Chen, Rongpeng Sun, Rui Yang, Rui Chen, Ruibin Zhou, Ruixu Zhang, Shaofeng Zhang, Sheng Xu, Shihao Chang, Shuaishuai Liu, Shulin Wang, SiQi Feng, Songjia Yuan, Songling Zhang, Tao Lang, Tianjiao Li, Tongkai Deng, Wei Li, Wei Wang, Weichao Zhang, Weigang Sun, Weixuan Ouyang, Wen Jiao, Wenxiang Sun, Wenzhi Jia, Wenzhuo Zhang, Xiang He, Xiangyu Ren, Xianshun Zhu, XiaoYing Guo, Xiaolong Li, Xiaoxue Ma, Xiaoyu Lu, Xican Feng, Xinhua Huang, Xinting Guan, Xinyu Li, Xirui Zhang, Xu Gao, Xudong Luo, Xun Qi, Xuxiang Chen, Yangkun Tao, Yangyu Xiao, Yanling Mai, Yantao Chen, Yanze Ding, Yao Yang, Yeting Song, YiFan Yang, Yifan Zhu, Yijiao Wu, Yinhe Liu, Yixian Yang, Yong Cai, Yuanjun Tu, Yuanlin Zhang, Yue Huang, Yufei Zhou, Yuhang Jiang, Yuhao Liu, Yuhong Hu, Yuhui Lin, Yujin Yang, Yun Wang, Yunhao Zhang, Yusong Wu, Zekun Zhang, Zelong Yu, Zhan Yang, Zhaoliang Zhao, Zhe Li, Zheng Huang, Zhenyu Liu, Zhiguang Xu, Zhijiang Kui, Zhiqing Zeng, Zhiyin Xiong, Zhiyuan Han, Zhuo Wu, Zifan Geng, Zigang Zhao, Zilong Tang, Ziyan Zhu, Ziyuan Zhu, Zonglei Xu, Zhijiang |
| author_facet | Tencent Hunyuan Team Liu, Ao Zhou, Botong Xu, Can Zhou, Chayse Zhang, ChenChen Xu, Chengcheng Wang, Chenhao Wu, Decheng Wu, Dengpeng Jiao, Dian Du, Dong Wang, Dong Zhang, Feng Lian, Fengzong Xu, Guanghui Zhang, Guanwei Wang, Hai Luo, Haipeng Hu, Han Xu, Huilin Wu, Jiajia Zhu, Jianchen Yan, Jianfeng Zhu, Jiaqi Zhang, Jihong Xue, Jinbao Xia, Jun Zheng, Junqiang Liu, Kai Zhang, Kai Zheng, Kai Li, Kejiao Wang, Keyao Jiang, Lan Liu, Lixin Wu, Lulu Huang, Mengyuan Yu, Peijie Wang, Peiqi Wang, Qian Xiang, Qianbiao Liu, Qibin Sun, Qingfeng Guo, Richard Xie, Ruobing Yang, Saiyong Chen, Shaohua Hu, Shihui Li, Shuai Li, Shuaipeng Chen, Shuang Zheng, Suncong Yang, Tao Zhang, Tian Yu, Tinghao Han, Weidong Liu, Weijie Zhou, Weijin Wang, Weikang Chen, Wesleye Feng, Xiao Ren, Xiaoqin Sun, Xingwu Kuang, Xiong Huang, Xuemeng Cao, Xun Chen, Yanfeng Du, Yang Yang, Zhen Tao, Yangyu Deng, Yaping Shen, Yi Hong, Yigeng Chen, Yiqi Huang, Yiqing Deng, Yuchi Mao, Yue Wang, Yulong Zeng, Yuyuan Xu, Zenan Kang, Zhanhui Zhao, Zhe Yan, ZhenXiang Fang, Zheng Hu, Zhichao Chen, Zhongzhi Li, Zhuoyu Li, Zongwei Yan, Alex Liang, Ande Liu, Baitong Pan, Beiping Xing, Bin Wu, Binghong Qu, Bingxin Ni, Bolin Wu, Boyu Li, Chen Jiang, Cheng Zhang, Cheng Liu, Chengjun Yang, Chengxu Xu, Chengzhong Wang, Chiyu Zha, Chong Yi, Daisy Wang, Di Lu, Fanyang Chen, Fei Liu, Feifei Zheng, Feng Yu, Guanghua Li, Guiyang Wang, Guohua Lin, Haisheng Liu, Han Wang, Han Fei, Hao Lu, Hao Jiang, Haoqing Sun, Haoran Zhu, Haotian Dai, Huangjin Chen, Huankui Feng, Huawen Cai, Huihui Peng, Huxin Lv, Jackson Shi, Jiacheng Bu, Jiahao Li, Jianbo Hu, Jianglu Guan, Jiangtao Xu, Jianing Cai, Jianwei Zhang, Jiarong Song, Jiawei Jiang, Jie Liu, Jie Yang, Jieneng Zhang, Jihong lv, Jin Zhao, Jing Li, Jinjian Liu, Jinxing Zhao, Jun Guo, Juntao Wang, Kai Wu, Kan Fu, Lei He, Lei Wang, Lei Liu, Li Dong, Liang Zhan, Liya Cheng, Long Xu, Long Zheng, Mao Liu, Meng Hu, Mengkang Chen, Nanli Chen, Peirui He, Peng Pan, Pengju Wei, Pengzhi Yang, Qi Yi, Qi Wang, Roberts Chen, Rongpeng Sun, Rui Yang, Rui Chen, Ruibin Zhou, Ruixu Zhang, Shaofeng Zhang, Sheng Xu, Shihao Chang, Shuaishuai Liu, Shulin Wang, SiQi Feng, Songjia Yuan, Songling Zhang, Tao Lang, Tianjiao Li, Tongkai Deng, Wei Li, Wei Wang, Weichao Zhang, Weigang Sun, Weixuan Ouyang, Wen Jiao, Wenxiang Sun, Wenzhi Jia, Wenzhuo Zhang, Xiang He, Xiangyu Ren, Xianshun Zhu, XiaoYing Guo, Xiaolong Li, Xiaoxue Ma, Xiaoyu Lu, Xican Feng, Xinhua Huang, Xinting Guan, Xinyu Li, Xirui Zhang, Xu Gao, Xudong Luo, Xun Qi, Xuxiang Chen, Yangkun Tao, Yangyu Xiao, Yanling Mai, Yantao Chen, Yanze Ding, Yao Yang, Yeting Song, YiFan Yang, Yifan Zhu, Yijiao Wu, Yinhe Liu, Yixian Yang, Yong Cai, Yuanjun Tu, Yuanlin Zhang, Yue Huang, Yufei Zhou, Yuhang Jiang, Yuhao Liu, Yuhong Hu, Yuhui Lin, Yujin Yang, Yun Wang, Yunhao Zhang, Yusong Wu, Zekun Zhang, Zelong Yu, Zhan Yang, Zhaoliang Zhao, Zhe Li, Zheng Huang, Zhenyu Liu, Zhiguang Xu, Zhijiang Kui, Zhiqing Zeng, Zhiyin Xiong, Zhiyuan Han, Zhuo Wu, Zifan Geng, Zigang Zhao, Zilong Tang, Ziyan Zhu, Ziyuan Zhu, Zonglei Xu, Zhijiang |
| contents | As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_15431 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought Tencent Hunyuan Team Liu, Ao Zhou, Botong Xu, Can Zhou, Chayse Zhang, ChenChen Xu, Chengcheng Wang, Chenhao Wu, Decheng Wu, Dengpeng Jiao, Dian Du, Dong Wang, Dong Zhang, Feng Lian, Fengzong Xu, Guanghui Zhang, Guanwei Wang, Hai Luo, Haipeng Hu, Han Xu, Huilin Wu, Jiajia Zhu, Jianchen Yan, Jianfeng Zhu, Jiaqi Zhang, Jihong Xue, Jinbao Xia, Jun Zheng, Junqiang Liu, Kai Zhang, Kai Zheng, Kai Li, Kejiao Wang, Keyao Jiang, Lan Liu, Lixin Wu, Lulu Huang, Mengyuan Yu, Peijie Wang, Peiqi Wang, Qian Xiang, Qianbiao Liu, Qibin Sun, Qingfeng Guo, Richard Xie, Ruobing Yang, Saiyong Chen, Shaohua Hu, Shihui Li, Shuai Li, Shuaipeng Chen, Shuang Zheng, Suncong Yang, Tao Zhang, Tian Yu, Tinghao Han, Weidong Liu, Weijie Zhou, Weijin Wang, Weikang Chen, Wesleye Feng, Xiao Ren, Xiaoqin Sun, Xingwu Kuang, Xiong Huang, Xuemeng Cao, Xun Chen, Yanfeng Du, Yang Yang, Zhen Tao, Yangyu Deng, Yaping Shen, Yi Hong, Yigeng Chen, Yiqi Huang, Yiqing Deng, Yuchi Mao, Yue Wang, Yulong Zeng, Yuyuan Xu, Zenan Kang, Zhanhui Zhao, Zhe Yan, ZhenXiang Fang, Zheng Hu, Zhichao Chen, Zhongzhi Li, Zhuoyu Li, Zongwei Yan, Alex Liang, Ande Liu, Baitong Pan, Beiping Xing, Bin Wu, Binghong Qu, Bingxin Ni, Bolin Wu, Boyu Li, Chen Jiang, Cheng Zhang, Cheng Liu, Chengjun Yang, Chengxu Xu, Chengzhong Wang, Chiyu Zha, Chong Yi, Daisy Wang, Di Lu, Fanyang Chen, Fei Liu, Feifei Zheng, Feng Yu, Guanghua Li, Guiyang Wang, Guohua Lin, Haisheng Liu, Han Wang, Han Fei, Hao Lu, Hao Jiang, Haoqing Sun, Haoran Zhu, Haotian Dai, Huangjin Chen, Huankui Feng, Huawen Cai, Huihui Peng, Huxin Lv, Jackson Shi, Jiacheng Bu, Jiahao Li, Jianbo Hu, Jianglu Guan, Jiangtao Xu, Jianing Cai, Jianwei Zhang, Jiarong Song, Jiawei Jiang, Jie Liu, Jie Yang, Jieneng Zhang, Jihong lv, Jin Zhao, Jing Li, Jinjian Liu, Jinxing Zhao, Jun Guo, Juntao Wang, Kai Wu, Kan Fu, Lei He, Lei Wang, Lei Liu, Li Dong, Liang Zhan, Liya Cheng, Long Xu, Long Zheng, Mao Liu, Meng Hu, Mengkang Chen, Nanli Chen, Peirui He, Peng Pan, Pengju Wei, Pengzhi Yang, Qi Yi, Qi Wang, Roberts Chen, Rongpeng Sun, Rui Yang, Rui Chen, Ruibin Zhou, Ruixu Zhang, Shaofeng Zhang, Sheng Xu, Shihao Chang, Shuaishuai Liu, Shulin Wang, SiQi Feng, Songjia Yuan, Songling Zhang, Tao Lang, Tianjiao Li, Tongkai Deng, Wei Li, Wei Wang, Weichao Zhang, Weigang Sun, Weixuan Ouyang, Wen Jiao, Wenxiang Sun, Wenzhi Jia, Wenzhuo Zhang, Xiang He, Xiangyu Ren, Xianshun Zhu, XiaoYing Guo, Xiaolong Li, Xiaoxue Ma, Xiaoyu Lu, Xican Feng, Xinhua Huang, Xinting Guan, Xinyu Li, Xirui Zhang, Xu Gao, Xudong Luo, Xun Qi, Xuxiang Chen, Yangkun Tao, Yangyu Xiao, Yanling Mai, Yantao Chen, Yanze Ding, Yao Yang, Yeting Song, YiFan Yang, Yifan Zhu, Yijiao Wu, Yinhe Liu, Yixian Yang, Yong Cai, Yuanjun Tu, Yuanlin Zhang, Yue Huang, Yufei Zhou, Yuhang Jiang, Yuhao Liu, Yuhong Hu, Yuhui Lin, Yujin Yang, Yun Wang, Yunhao Zhang, Yusong Wu, Zekun Zhang, Zelong Yu, Zhan Yang, Zhaoliang Zhao, Zhe Li, Zheng Huang, Zhenyu Liu, Zhiguang Xu, Zhijiang Kui, Zhiqing Zeng, Zhiyin Xiong, Zhiyuan Han, Zhuo Wu, Zifan Geng, Zigang Zhao, Zilong Tang, Ziyan Zhu, Ziyuan Zhu, Zonglei Xu, Zhijiang Computation and Language As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models. |
| title | Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2505.15431 |