DeepSeek-V4-Pro

DeepSeek-AI, Anyi Xu, Bang Lin, Bing Xue, Bing-Li Wang, Bin Xu, Bo Wu, Bowei Zhang, Chao Lin, Chengyao Dong, Chen Ling, Chengda Lu, Chen Zhao, Chengqi Deng, C. Hou, Chen Xu, Chenze Shao, Chong Ruan, Conner Sun, Damai Dai, Daya Guo, De-Bin Yang, Deli Chen, Dong-Hui Li, Dong-Li Ji, Erhang Li, Fangchen Wei, Fangyun Lin, Fang Yuan, Fei Xia, Fucong Dai, Guangbo Hao, Guanting Chen, Guo Cao, Guo-Hui Meng, Guowei Li, Han Yu, Han Zhang, Hanwei Xu, Hao Li, Hao Liang, Haoling Zhang, Haoming Luo, Haoran Wei, Hao Yuan, Haowei Zhang, Haowen Luo, Hao Chen, Haozhe Ji, Heng Zhang, Honghui Ding, Hongxuan Tang, Huanqi Cao, Huazuo Gao, Hui Qu, Hui Zeng, J. Yang, JQ Zhu, Jianming Luo, Jia Song, Jia Yu, Jialiang Huang, Jialu Cai, Jian Liang, Jiangting Zhou, Jiasheng Ye, Jiashi Li, Jiaxin Xu, Jiewen Hu, Jieyu Yang, Jin Chen, Jin Yan, JingChang Chen, Jing Zhou, Jing Xiang, Jingyang Yuan, Jing Cheng, Jingzi Zhou, Jinhua Zhu, Jiping Yu, Joseph Sun, Junliang Ran, Jun Jiang, Junjie Qiu, Junlong Li, Junming Zheng, Jun-Mei Song, Kai Dong, Kaige Gao, Kang Guan, Kexing Zhou, Kezhao Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Wang, Leyi Xia, Li Zhang, Liang Zhao, Lihua Guo, Lin Luo, Lin Ma, Linyan Zhu, Litong Wang, Liyu Cai, Liyue Zhang, Longhao Chen, Mingxi Di, My Xu, Max Mei, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Mi Zhou, Min Han, Ning Wang, Pan Huang, Pan Wang, Peixin Cong, Peiyi Wang, Peng Zhang, Qiancheng Wang, Qihao Zhu, Qingyang Li, Qinyu Chen, Qiushi Du, Qi Jiang, Rui Tian, R. Xu, Ruijie Lu, Ruiling Xu, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. Chen, Runqiu Yin, Runxin Xu, Ruo-Han Shen, Ruoyu Zhang, Ruyi Chen, Sh. Liu, Shanghao Lu, Shangmian Sun, Shangyan Zhou, Shan-Shan Chen, Shaofei Cai, Shao Nie, Shao-Ping Wu, Shaoyuan Chen, Shengding Hu, Sheng Liu, Shiqiang Hu, Shirong Ma, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, Shuying Yu, Songyang Zhou, Tao Ni, Tao Yun, Tian Jin, Tianhong Pei, Tian Ye, Tianle Lin, Tian Ji, Tianyi Cui, Tianyuan Yue, Tingting Yu, Tu-Tu Wang, W. Zhang, Wl Xiao, Wangding Zeng, Wei An, Weilin Zhao, Wen Liu, W. Liang, Wenjie Pang, Wenjing Luo, Wenjin Yao, Wenjun Gao, Wenkai Yang, Wen-Wen Huang, Wenqing Hou, Wentao Zhang, Wenting Ma, Xi Gao, Xiang He, Xiang Wang, Xianzu Wang, Xiao Bi, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaokang Zhang, X. Nie, Xiaowen Sun, Xiaoxiang Wang, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xingchen Liu, Xingkai Yu, Xingyou Li, Xinyu Yang, Xinyu Zhang, Xu Chen, Xuanyu Wang, Xuecheng Su, Xueyin Chen, Xuheng Lin, Xu Fu, Yc Yan, YQ Wang, Yw Ma, Yanfen Luo, Yang Zhang, Yanhong Xu, Yanru Ma, Yanwen Huang, Yao Li, Yao Xu, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Qian, Yingxia Shao, Yi Yu, Yichao Zhang, Yifan Ding, Yifan Shi, Yijia Wu, Yi-Yu Xiong, Yi Ma, Ying He, Ying Tang, Ying Zhou, Yingjia Luo, Yinmin Zhong, Y. Piao, Yisong Wang, Yixiang Zhang, Yixiao Chen, Yixuan Tan, Yixuan Wei, Yiyang Ma, Yiyuan Liu, Yong Yang, Yongqiang Guo, Yongtong Wu, Yu Wu, Yukun Li, Yuan Cheng, Yuan Ou, Yuanfan Xu, Yuanhao Li, Yuduan Wang, Yuehan Yang, Yue Xu, Yuhan Wu, Yu Meng, Yu Zou, Y. Zha, Yunfan Xiong, Yupeng Chen, Yuping Lin, Yu Cao, Yuqian Wang, Yushun Zhang, Yuting Yan, Yutong Lin, Yuxian Gu, Yu-Wei Luo, Yu-mei You, Yuxuan Liu, Yuxuan Zhou, Yuyang Zhou, Yuzhen Huang, Zhanghua Wu, Ze Wang, Ze Zhao, Z. Ren, Zekai Zhang, Zhangli Sha, Zhe Fu, Zhe Ju, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhe Gao, Zhewen Hao, Zhibin Gou, Zhicheng Ma, Zhigang Yan, Zhihong Shao, Zhixiang Huang, Zhixuan Chen, Zhiyu Wu, Zhizhou Ren, Zhongyu Wu, Zhuoshu Li, Zhuping Zhang, Zian Xu, Zihao Wang, Z. Qu, Zihui Gu, Zijia Zhu, Zi-Long Li, Zipeng Zhang, Ziwei Xie, Ziyi Gao, Ziyi Wan, Zizheng Pan, Zong-xiang Yao

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models -- DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) -- both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: (1) a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency; (2) Manifold-Constrained Hyper-Connections (mHC) that enhance conventional residual connections; (3) and the Muon optimizer for faster convergence and greater training stability. We pre-train both models on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline that unlocks and further enhances their capabilities. DeepSeek-V4-Pro-Max, the maximum reasoning effort mode of DeepSeek-V4-Pro, redefines the state-of-the-art for open models, outperforming its predecessors in core tasks. Meanwhile, DeepSeek-V4 series are highly efficient in long-context scenarios. In the one-million-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. This enables us to routinely support one-million-token contexts, thereby making long-horizon tasks and further test-time scaling more feasible. The model checkpoints are available at https://huggingface.co/collections/deepseek-ai/deepseek-v4.

Eduardo C. Garrido-Merchán, 'Alvaro L'opez L'opez

Generative artificial intelligence, and large language models in particular, have emerged as one of the most transformative paradigms in modern computer science. This automated survey provides an accessible treatment of the field as of early 2026, with a strong focus on the leading model families, deployment protocols, and real-world applications. The core of the survey is devoted to a detailed comparative analysis of the frontier large language models, with particular emphasis on open-weight systems: DeepSeek-V3, DeepSeek-R1, DeepSeek-V3.2, and the forthcoming DeepSeek V4; the Qwen 3 and Qwen 3.5 series; GLM-5; Kimi K2.5; MiniMax M2.5; LLaMA 4; Mistral Large 3; Gemma 3; and Phi-4, alongside proprietary systems including GPT-5.4, Gemini 3.1 Pro, Grok 4.20, and Claude Opus 4.6. For each model, we describe the architectural innovations, training regimes, and empirical performance on current benchmarks and the Chatbot Arena leaderboard. The survey further covers deployment protocols including Retrieval-Augmented Generation, the Model Context Protocol, the Agent-to-Agent protocol, function calling standards, and serving frameworks. We present an extensive review of real-world applications across fifteen industry sectors, from financial services and legal technology to tourism and agriculture, supported by empirical evidence and case studies. This work has been generated by Claude Opus 4.6 (Anthropic) under the supervision and editorial review of the human authors, with the goal of producing updated editions approximately every six months.

Olaf Yunus Laitinen Imanov

Sequential fine-tuning of Large Language Models (LLMs) adaptation to target tasks often triggers catastrophic forgetting, where the acquisition of novel target skills degrades ancestral capabilities. This paper presents a systematic comparative study of catastrophic forgetting across twenty premier models representing the state-of-the-art in mid-2026. We categorize our investigation into two primary research lines: (i) a behavioral and semantic output drift analysis of ten leading closed-source models (including Claude Fable 5, GPT-5.5 High, and Gemini 3.5 Flash), and (ii) a deep mechanistic interpretation of ten prominent open-weight architectures (such as DeepSeek-V4-Pro, Llama 4 Maverick, and Qwen 3.6-27B). Through weight-space trajectory tracking, Centered Kernel Alignment (CKA), and routing gate drift calculations in Mixture-of-Experts (MoE) layers, we localize the neural circuits highly susceptible to parameter overwriting. Our findings indicate that early-layer attention heads exhibit systemic entropic dispersion, while mid-to-deep feed-forward networks (or sparse expert blocks) suffer localized representation collapse. Informed by these insights, we introduce Low-Rank Circuit Projection (LRCP), a subspace-regularized training intervention. Empirical evaluations show that LRCP successfully mitigates up to 94.2% of ancestral capabilities in open-weight configurations and matches the adaptation velocity of standard PEFT baselines.

Shanshan Wang, Sizhen Zhu, Rende Li

Tool-calling LLM agents are vulnerable to indirect prompt injection: externally retrieved data can redirect tool calls without system-prompt access, and prompt-level defences leave three harm classes undefended (path traversal, user-guided exfiltration, high-frequency tool abuse). We present a Policy Enforcement Point (PEP) that intercepts at the tool-call boundary with declarative rules over a cross-step information-flow label system (source integrity, data sensitivity) and a synchronous SHA-256 hash-chained audit log. On a controlled dataset across four attack classes, the full system cuts the attack success rate (ASR) from 40.0% to 5.0% (deepseek-v4-pro, five repeats) versus 35.0% for the strongest prompt-only baseline; disabling cross-step label propagation raises the call-level false-negative rate by 26.4 points. The 30.0% task-level false-positive rate is dominated by by-design least-privilege capability-token denials, not rule false positives—an expanded 30-task benign set yields 0/30 rule false positives under scripted isolation. A conservative-DS mitigation (intent-taint) closes the constructed denied-read reconstruction blind-spot variant (ASR 100% to 0%) at no cost on standard workflows. The audit log detects all three tested tamper classes; the in-process enforcement overhead is sub-millisecond per call. Across four further backends, ASR drops under the full system, though LLaMA-3.3-70B retains 16.7% (a rule-coverage gap). A preliminary run over a real MCP stdio transport (an official filesystem server) shows the mechanism operates at a real boundary with a sub-millisecond execution-path increment. We frame these as mechanism-coverage evidence on a controlled benchmark, not a deployability claim for production MCP workloads. Code, data, and metrics are openly available in the replication repository.

Mao Zheng, Zhengnan Li, Tao Chen, Bo Lv, Ming Sun, Mingyang Song, Jinlong Song, Hong Huang, Decheng Wu, Hai Wang, Yifan Song, Yanfeng Chen, Guanwei Zhang, Guanghua Yu, Yi Su, Hong Liu, Jinxiang Ou, Keyao Wang, Wei-Lin Chen, Haozhao Kuang, Kai Wang, Nuo Chen, Zihao Zheng, Chenhao Wang, Bingfa Xing, Chengcheng Xu, Tinghao Yu, Binghong Wu, Longfei Xu, Jiacheng Shi, Yunhao Wang, Bai-Tao Chen, Lei Zhang, Qi Yang, Zhaoren Wu, Jiachen Li, Lan Jiang, Lanrui Wang, Kai Zhang, Shuaipeng Li, Zhong Chen, Weixuan Sun, Jiaqi Zhu, Andi Wang, Wei Li, Jun Xia, Weidong Han, Wu-Ming Yang, Li Hui, Lu Jia, Jiajia Wu, Xin Zhou, Tianxian Fei

Hy-MT2 is a family of fast-thinking multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall. Moreover, when paired with AngelSlim's 1.25-bit extreme quantization for on-device deployment, the lightweight 1.8B model requires only 440 MB of storage and achieves a 1.5x inference speedup.