Authors
DeepSeek-AI:Anyi XuB. LiBangcai LinBing XueBingCheng XianBingzheng XuBochao WuBowei ZhangBoyi DengC. C. YuChao JinChaofan LinChen DongChenbing WangChenfan FengChengda LuChenggang ZhaoChengqi DengChengyuan ZhangChenhao XuChenqi ZhaoChenze ShaoChuhao WangChuqi ZhangDamai DaiDejian YangDeli ChenDi HuangDi WuDonghao LiErhang LiEric FuF. ZhouFangwei ZhouFangyun LinFangzhou YuanFeiyu XiaFucong DaiGuangbo HaoGuanglin LiGuanting ChenGuoai CaoGuofan FanGuolai MengGuowei LiHaichuan ZhangHaiyang MaHaiyang ShenHan LiHan YuHan ZhangHangyuan DengHanwei XuHanxiang XuHanxun ZhongHao GuoHao JiangHao LiHao QinHaodong WenHaofen LiangHaofeng HuangHaohua LiuHaoling ZhangHaoming LuoHaoran YangHaotian XuHaotian YuanHaoting HuangHaowen LuoHaoyang CaiHaoyu ChenHaozhe JiHengran ZhangHengrui WangHengxu WuHonghui DingHongxuan TangHuadong WangHuanqi CaoHuazuo GaoHui QuHui ZengJ. YangJ. H. JinJ. H. ZhangJ. X. ZouJia YuJiahui ZhouJiajun ChenJialiang HuangJialin ZhaoJiamin TangJian ZhouJianan TongJianwen LiJiaqi ZhuJiarui WangJiasheng YeJiashi LiJiaxin XuJiaying DingJibai LuJiewen HuJin YanJincheng ZhaiJingchang ChenJingcheng HuJingli ZhouJingsheng XuJingting XiangJingyan YunJingyang YuanJingyuan ChengJinhua ZhuJinpeng WangJinyi ChenJinyi HuJiping YuJueliang GuoJunbo PeiJunbo SunJunguang JiangJunjie QiuJunkang ZhouJunqi LiuJunren LiJunxian LiJunxiao SongJunyi GuoKai DongKaifeng ChenKaige GaoKang GuanKangdong YuanKe HongKe XuKefan ZhaoKexin JiKexin ZhangKexing ZhouKuai YuLan ZhangLean WangLecong ZhangLei WangLetian GaoLiang ZhaoLiansheng XuLihua GuoLingxiao LuoLingyue FuLitao DengLitong WangLiyue ZhangLonghao ChenLu ChenLuotian HuangLuyao MaLuyao WangM. S. DiMax MeiMenghao YeMiao CuiMingchuan ZhangMinghua ZhangMinghui TangMingjing ZhangMingqi WeiMingshu ChenMingxing LiuMingxu ZhouMingyu XuMingyu YangMingze WangMuyang ChenNi ShentuNing WangNiufang NingPanpan HuangPeixin CongPeiyi WangPeiyuan XinPengfei RenPengfei YanPengle ZhangQi KangQi TangQiancheng WangQiang LiQihao ZhuQingyang LiQinyu ChenQiushi DuQizhou GuoRongxian XuRui DingRui HuRui TianRui YuRuidong ZhuRuifan XuRuihan YangRuihang XiaRuijie LuRuilin GengRuipeng HongRuiqi GeRuisong ZhangRuize SunRuizhe PanRunji WangRunqian ChenRunxin XuRuohong TianRuomeng ShenRuoyu ZhangRyan X.S. H. LiuShanghao LuShangyan ZhouShanhuang ChenShaofei CaiShaoheng NieShaoyuan ChenShengding HuShengkai LinShengwen RanShengyu LiuShengyuan JiaShi BaiShi FengShicheng XuShichun LiuShiqiang HuShirong MaShiyu WangShiyuan FengShufan GongShuhan LinShuiping YuShunfeng ZhouShuo YangShuomeng WangShuting GuoShuting PanShuying YuSinuo CaoSiyi LinSizhe ChenSongyang ChenSongyang ZhouTao NiTao YunTian JinTian PeiTian YeTianle LinTianran JiTianyi CuiTianyuan YueTingting YuTongrui XiongWangding ZengWei LiuWei ZhangWeibin XuWeihao ZengWeilin ZhaoWen LiuWenfeng LiangWenjie PangWenjing LuoWenjing YaoWenjun GaoWenkai ShaoWenkai YangWenli ZhangWenlu WangWenlve HuangWenqian YanWentao ZhangXi GaoXiang HeXiang LiXiangli LiXiangwen WangXiangying ZhangXiankui WeiXiao BiXiaodong LiuXiaohan WangXiaojian QuXiaokang ChenXiaokang ZhangXiaotao NieXiaoyao ZouXiaoyuan LiXicheng GuoXieting ChuXin ChengXin LiuXin XieXinbo XuXingchao LiuXingchen LiuXingkai YuXingyou LiXintong YaoXinyang ChenXinyong JiangXinyu YangXinyu YangXu ChenXuanyu WangXubei ZhongXuecheng SuXuejie LiuXuheng LinXujie FanXuncheng ZhaoXuwei FuY. C. YanY. H. JiangY. T. WuY. W. M.Y. Z. WangYafei GaoYang YangYang ZhangYanru MaYanwen HuangYao LiYao LiYao MengYao ZhaoYaofeng SunYaohui WangYaoyang YeYehang YinYexinrui WuYi QianYi TaoYi YuYichao ZhangYichen JiangYicheng WangYifan DingYifan ShiYifeng PengYifeng ZhaiYijia WuYiliang XiongYilun WangYing HeYing ZhouYingjia LuoYinmin ZhongYiping WangYisong WangYixiang ZhangYixiao ChenYixuan TanYixuan WeiYiyang MaYiyao YangYiyuan LiuYizai CaiYizhen WeiYizhi WangYonglun YangYongqi ZhuoYongqiang GuoYongtong WuYu WuYu ZhangYuan BianYuan ChengYuan OuYuan SunYuanfan XuYuanhang SunYuanhao LiYuchen LiuYuchen YaoYudong HanYuduan WangYuhan WuYuhao MengYuheng ZouYuKun LiYunchuan WangYunfan XiaoYunfan XiongYupeng ChenYuqian CaoYuqian WangYuqing ChenYushun ZhangYutong LinYuwei XiaoYuxian GuYuxiang ChenYuxiang HuangYuxiang LuoYuxiang YouYuxin ChenYuxin XiangYuxuan LiuYuxuan ZhouYuyang ZhouYuzhe GuoYuzhen HuangYuzhuo BaiZ. Y. Z.Zanlin NiZehao WangZehua ZhaoZehui RenZejun ZhaoZhangli ShaZhanying WangZhaochen ZhangZhaoshuai DuZhe FuZhean XuZhenda XieZheng LiuZhengyan ZhangZhenhua DongZhewen HaoZhibang WangZhibin GouZhicheng MaZhihao LiZhihong ShaoZhihuan HuangZhijie LiZhirui LuZhixian HuangZhixuan ChenZhixuan ChenZhixuan PanZhiyu WuZhizhou RenZhu HeZhuoshu LiZhuping ZhangZian XuZihao WangZihui GuZijia ZhuZili ZhangZilin LiZilong HouZilong LyuZiqiao WangZiwei XieZiya ZhangZiyi GaoZizheng PanZonglin LiZongqing YaoZui ChenZuofan WuChenchen LingChengyu HouChong ChenD. LiDi QiDongjie JiFang WeiFanyi XiaFei XieFeiyi TanHailong GuoHaiyan ZhaiHui ZhouHuihui TanHuijie LiJia LuoJia SongJialu CaiJian LiangJiangting ZhouJiaqi GaoJiayi ShaoJie ChenJieyu YangJin ChenJingde ZhangJingzi ZhouJinqian WangJinyang LiuJinZhao SunJunhua LingJunmin ZhengKaicheng YangKe XuLe SuLeyi XiaLiangfeng DingLin ZhuoLinwang MaLinyan ZhuLiyu CaiLuqi YaoM. K. ZhangMeng LiMiao LinMiaojun WangMin ZhangMingming LiMingming WangMingze YinMinmin HanNan CaoNing WangNingxin MaPanpan WangPeihan LinPeng SunPeng ZhangQian YingQiang XiangQiao WangQingmiao MaoQiwei JiangRongli JinRuyi ChenSha TaoShangmian SunShaoqing WuShichao ZouSi LeiTianyang ZhangTianyu SunTingting YinW. L. XiaoWei AnWei LiWei WangWeiwei LinWenqing HouX. LinXiangfei MengXianzhu HuangXiao PengXiaoqian LiXiaoting ZhangXiaowen SunXiaoxiang WangXiaoyu YeXinrou ZhangXinyu ZhangXue CaoXueyin ChenYanan ZhouYanhong XuYao XiaYao XuYi ShaoYihong ZhangYiling MaYing TangYining LouYiru ChenYishi PiaoYixuan ChenYong XiongYuchen XuanYuehan YangYuer XuYukun ZhaYunxian MaYuping LinYuting YanYutong XieYuwen ShengYuxuan ZhuZekai ZhangZhe JuZhenzhen LinZheren GaoZheyang SunZhigang YanZhongyu WuZi WangZihua QuZiling YanZiyi Wan
Abstract
The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.