Cover Image

Saved in:
Bibliographic Details
Main Authors: DeepSeek-AI, Liu, Aixin, Mei, Aoxue, Lin, Bangcai, Xue, Bing, Wang, Bingxuan, Xu, Bingzheng, Wu, Bochao, Zhang, Bowei, Lin, Chaofan, Dong, Chen, Lu, Chengda, Zhao, Chenggang, Deng, Chengqi, Xu, Chenhao, Ruan, Chong, Dai, Damai, Guo, Daya, Yang, Dejian, Chen, Deli, Li, Erhang, Zhou, Fangqi, Lin, Fangyun, Dai, Fucong, Hao, Guangbo, Chen, Guanting, Li, Guowei, Zhang, H., Xu, Hanwei, Li, Hao, Liang, Haofen, Wei, Haoran, Zhang, Haowei, Luo, Haowen, Ji, Haozhe, Ding, Honghui, Tang, Hongxuan, Cao, Huanqi, Gao, Huazuo, Qu, Hui, Zeng, Hui, Huang, Jialiang, Li, Jiashi, Xu, Jiaxin, Hu, Jiewen, Chen, Jingchang, Xiang, Jingting, Yuan, Jingyang, Cheng, Jingyuan, Zhu, Jinhua, Ran, Jun, Jiang, Junguang, Qiu, Junjie, Li, Junlong, Song, Junxiao, Dong, Kai, Gao, Kaige, Guan, Kang, Huang, Kexin, Zhou, Kexing, Huang, Kezhao, Yu, Kuai, Wang, Lean, Zhang, Lecong, Wang, Lei, Zhao, Liang, Yin, Liangsheng, Guo, Lihua, Luo, Lingxiao, Ma, Linwang, Wang, Litong, Zhang, Liyue, Di, M. S., Xu, M. Y, Zhang, Mingchuan, Zhang, Minghua, Tang, Minghui, Zhou, Mingxu, Huang, Panpan, Cong, Peixin, Wang, Peiyi, Wang, Qiancheng, Zhu, Qihao, Li, Qingyang, Chen, Qinyu, Du, Qiushi, Xu, Ruiling, Ge, Ruiqi, Zhang, Ruisong, Pan, Ruizhe, Wang, Runji, Yin, Runqiu, Xu, Runxin, Shen, Ruomeng, Zhang, Ruoyu, Liu, S. H., Lu, Shanghao, Zhou, Shangyan, Chen, Shanhuang, Cai, Shaofei, Chen, Shaoyuan, Hu, Shengding, Liu, Shengyu, Hu, Shiqiang, Ma, Shirong, Wang, Shiyu, Yu, Shuiping, Zhou, Shunfeng, Pan, Shuting, Zhou, Songyang, Ni, Tao, Yun, Tao, Pei, Tian, Ye, Tian, Yue, Tianyuan, Zeng, Wangding, Liu, Wen, Liang, Wenfeng, Pang, Wenjie, Luo, Wenjing, Gao, Wenjun, Zhang, Wentao, Gao, Xi, Wang, Xiangwen, Bi, Xiao, Liu, Xiaodong, Wang, Xiaohan, Chen, Xiaokang, Zhang, Xiaokang, Nie, Xiaotao, Cheng, Xin, Liu, Xin, Xie, Xin, Liu, Xingchao, Yu, Xingkai, Li, Xingyou, Yang, Xinyu, Li, Xinyuan, Chen, Xu, Su, Xuecheng, Pan, Xuehai, Lin, Xuheng, Fu, Xuwei, Wang, Y. Q., Zhang, Yang, Xu, Yanhong, Ma, Yanru, Li, Yao, Zhao, Yao, Sun, Yaofeng, Wang, Yaohui, Qian, Yi, Yu, Yi, Zhang, Yichao, Ding, Yifan, Shi, Yifan, Xiong, Yiliang, He, Ying, Zhou, Ying, Zhong, Yinmin, Piao, Yishi, Wang, Yisong, Chen, Yixiao, Tan, Yixuan, Wei, Yixuan, Ma, Yiyang, Liu, Yiyuan, Yang, Yonglun, Guo, Yongqiang, Wu, Yongtong, Wu, Yu, Cheng, Yuan, Ou, Yuan, Xu, Yuanfan, Wang, Yuduan, Gong, Yue, Wu, Yuhan, Zou, Yuheng, Li, Yukun, Xiong, Yunfan, Luo, Yuxiang, You, Yuxiang, Liu, Yuxuan, Zhou, Yuyang, Wu, Z. F., Ren, Z. Z., Zhao, Zehua, Ren, Zehui, Sha, Zhangli, Fu, Zhe, Xu, Zhean, Xie, Zhenda, Zhang, Zhengyan, Hao, Zhewen, Gou, Zhibin, Ma, Zhicheng, Yan, Zhigang, Shao, Zhihong, Huang, Zhixian, Wu, Zhiyu, Li, Zhuoshu, Zhang, Zhuping, Xu, Zian, Wang, Zihao, Gu, Zihui, Zhu, Zijia, Li, Zilin, Zhang, Zipeng, Xie, Ziwei, Gao, Ziyi, Pan, Zizheng, Yao, Zongqing, Feng, Bei, Li, Hui, Cai, J. L., Ni, Jiaqi, Xu, Lei, Li, Meng, Tian, Ning, Chen, R. J., Jin, R. L., Li, S. S., Zhou, Shuang, Sun, Tianyu, Li, X. Q., Jin, Xiangyue, Shen, Xiaojin, Chen, Xiaosha, Song, Xinnan, Zhou, Xinyi, Zhu, Y. X., Huang, Yanping, Li, Yaohui, Zheng, Yi, Zhu, Yuchen, Ma, Yunxian, Huang, Zhen, Xu, Zhipeng, Zhang, Zhongyu, Ji, Dongjie, Liang, Jian, Guo, Jianzhong, Chen, Jin, Xia, Leyi, Wang, Miaojun, Li, Mingming, Zhang, Peng, Chen, Ruyi, Sun, Shangmian, Wu, Shaoqing, Ye, Shengfeng, Wang, T., Xiao, W. L., An, Wei, Wang, Xianzu, Sun, Xiaowen, Wang, Xiaoxiang, Tang, Ying, Zha, Yukun, Zhang, Zekai, Ju, Zhe, Zhang, Zhen, Qu, Zihua
Format: Preprint
Published: 2025
Subjects:
Computation and Language
Online Access:https://arxiv.org/abs/2512.02556
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments.

Similar Items