21 lines
16 KiB
Plaintext
21 lines
16 KiB
Plaintext
=== https://huggingface.co/deepseek-ai (108097 chars, 0 arxiv refs)
|
||
TEXT: deepseek-ai (DeepSeek) Hugging Face Models Datasets Spaces Buckets new Docs Enterprise Pricing Website Tasks HuggingChat Collections Languages Organizations Community Blog Posts Daily Papers Hardware Learn Discord Forum GitHub Solutions Team & Enterprise Hugging Face PRO Enterprise Support Inference Providers Inference Endpoints Storage Buckets Log In Sign Up DeepSeek company Verified https://www.deepseek.com/ deepseek_ai deepseek-ai Activity Feed Follow 146,820 DeepSeek (深度求索), founded in 2023, is a Chinese company dedicated to making AGI a reality.</p>\n<p>Unravel the mystery of AGI with curiosity. Answer the essential question with long-termism.</p>\n","classNames":"hf-sanitized hf-sanitized-0NGt-rGwm-LYepq2rdiP6"},"users":[{"_id":"652faf2b1ad13fee8c4861de","avatarUrl":"/avatars/c3862c9cc8f1e645bb13f516c57e4e0e.svg","isPro":false,"fullname":"DeepSeek","user":"deepseek-admin","type":"user"},{"_id":"63cd76b4374057a338e8e703","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63cd76b4374057a338e8e703/i4Qk5-0aYx3oRhC8b50aJ.jpeg","isPro":false,"fullname":"zhuqihao","user":"zqh11","type":"user"},{"_id":"6539f6ea26df26ecd1393c37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6539f6ea26df26ecd1393c37/9VJusLLAiLhUxfBAFggpF.jpeg","isPro":false,"fullname":"Zhenda Xie","user":"zdaxie","type":"user"},{"_id":"6225bb44c6e650de3a65dbaa","avatarUrl":"/avatars/99c99ced2461978df572c27c1b3a4904.svg","isPro":false,"fullname":"DejianYang","user":"DejianYang","type":"user"},{"_id":"6545c3d0cf50edb69fbf7a67","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6545c3d0cf50edb69fbf7a67/lCfz3ILf5dHXf_lM2KBST.jpeg","isPro":false,"fullname":"sule","user":"sule91","type":"user"},{"_id":"6482e57a04f67f5f6056a61b","avatarUrl":"/avatars/b26faf19ba1493b91102ac7978ab3230.svg","isPro":false,"fullname":"Shirong Ma","user":"msr2000","type":"user"},{"_id":"6566a5af5ed6ca5f18ea44b6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6566a5af5ed6ca5f18ea44b6/8VAU2ImBFj5HOasuu9_W8.png","isPro":false,"fullname":"Freja","user":"Freja71122","type":"user"},{"_id":"6568596e6fcc82e5e8f66ffa","avatarUrl":"/avatars/12ec4239eae394627fce933ca18cf620.svg","isPro":false,"fullname":"Auden Wu","user":"Auden567","type":"user"},{"_id":"659384dec0b1372b2e3c184d","avatarUrl":"/avatars/fdbff8012fe80d5f2e861cbfa2675125.svg","isPro":false,"fullname":"wangding zeng","user":"zwd973-deepseek","type":"user"},{"_id":"659389f8de82e1ef7b9a8b13
|
||
=== https://github.com/deepseek-ai (4557 chars, 0 arxiv refs)
|
||
TEXT: DeepSeek · GitHub Skip to content Navigation Menu Sign in Appearance settings deepseek-ai Platform AI CODE CREATION GitHub Copilot Write better code with AI GitHub Copilot app Direct agents from issue to merge MCP Registry Integrate external tools DEVELOPER WORKFLOWS Actions Automate any workflow Codespaces Instant dev environments Issues Plan and track work Code Review Manage code changes Code Quality Enforce quality at merge APPLICATION SECURITY GitHub Advanced Security Find and fix vulnerabilities Code security Secure your code as you build Secret protection Stop leaks before they start EXPLORE Why GitHub Documentation Blog Changelog Marketplace View all features Solutions BY COMPANY SIZE Enterprises Small and medium teams Startups Nonprofits BY USE CASE App Modernization DevSecOps DevOps CI/CD View all use cases BY INDUSTRY Healthcare Financial services Manufacturing Government View all industries View all solutions Resources EXPLORE BY TOPIC AI Software Development DevOps Security View all topics EXPLORE BY TYPE Customer stories Events & webinars Ebooks & reports Business insights GitHub Skills SUPPORT & SERVICES Documentation Customer support Community forum Trust center Partners View all resources Open Source COMMUNITY GitHub Sponsors Fund open source developers PROGRAMS Security Lab Maintainer Community GitHub Stars Archive Program REPOSITORIES Topics Trending Collections Enterprise ENTERPRISE SOLUTIONS Enterprise platform AI-powered developer platform AVAILABLE ADD-ONS GitHub Advanced Security Enterprise-grade security features Copilot for Business Enterprise-grade AI features Premium Support Enterprise-grade 24/7 support Pricing Search / Sign in Sign up Appearance settings You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session. Dismiss alert {{ message }} DeepSeek 106k followers https://www.deepseek.com/ service@deepseek.com Overview Repositories Projects Packages People More Overview Repositories Projects Packages People Pinned Loading deepseek-harness deepseek-harness Public DeepSeek Harness: Everything is a Plugin. TypeScript 231k 27.6k Repositories --> Loading Type Select type All Public Sources Forks Archived Mirrors Templates Language Select language All C++ Cuda Makefile Python Rust TypeScript Sort Select order Last updated Name Stars Showing 10 of 39 repositories deepsee
|
||
=== https://huggingface.co/papers?q=DeepSeek (3594 chars, 0 arxiv refs)
|
||
TEXT: Daily Papers - Hugging Face Hugging Face Models Datasets Spaces Buckets new Docs Enterprise Pricing Website Tasks HuggingChat Collections Languages Organizations Community Blog Posts Daily Papers Hardware Learn Discord Forum GitHub Solutions Team & Enterprise Hugging Face PRO Enterprise Support Inference Providers Inference Endpoints Storage Buckets Log In Sign Up new Get trending papers in your email inbox once a day! Get trending papers in your email inbox! Subscribe Daily Papers by AK and the research community Daily Weekly Monthly Sep 18 Submitted by taesiri 120 DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression DeepSeek 3 Submitted by ZhaoHaoyuu 93 Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model · 13 authors 22 4 Submitted by kzhao5 92 When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation Microsoft 7 5 Submitted by Sensen02 88 SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness NVIDIA 2.42k 3 Submitted by Vfrz 69 An Empirical Study of Harness Design for Coding Agents Zoom Communications 3 Submitted by taesiri 53 JEPA-Anything: Learning Predictive Models across Different Worlds · 13 authors 64 4 Submitted by zorik 44 Verifiable Social Reasoning for LLM Assistants Google 3 Submitted by augustinLib 40 Self-Evolving Search Index · 8 authors 3 Submitted by leonliuzx 40 RiskChainBench: A Benchmark for Obfuscated Platform Message Restoration and Evidence-Grounded Web Investigation · 13 authors 1 4 Submitted by LZXzju 37 RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning · 11 authors 3 Submitted by LZXzju 37 Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents · 10 authors 13 2 Submitted by Longin-Yu 34 WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing Tencent 106 2 Submitted by taesiri 34 Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation · 11 authors 2 Submitted by MrBean2024 33 VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control DaLian University of Technology 12 2 Submitted by junshim 31 When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models Microsoft 2 Submitted by iammobina 29 Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling iowa state university 4 Submitted by xiuyuz 29 What
|
||
=== https://deepseekai.guide/news/deepseek-research-papers/ (24253 chars, 8 arxiv refs)
|
||
id: 2401.02954
|
||
id: 2401.06066
|
||
id: 2401.14196
|
||
id: 2402.03300
|
||
id: 2405.04434
|
||
id: 2412.19437
|
||
id: 2501.12948
|
||
id: 2512.02556
|
||
TEXT: DeepSeek Research Papers: V4 to V2 Reading Guide Skip to content No results Home Models Tutorials Comparisons Guides Reviews API Tools About DeepSeek AI Guide Home Models Tutorials Comparisons Guides Reviews API Tools About DeepSeek AI Guide Search Menu A Guide to DeepSeek Research Papers From V2 to V4 Map every major DeepSeek research paper from V2 to the V4 technical report, with key contributions and benchmarks. Start reading today. April 24, 2026 News Quick facts Quick facts Topic A Guide to DeepSeek Research Papers From V2 to V4 Format News brief Models discussed V4-Pro, V4-Flash, V4, V3.2 Reading time 11 min read (~2,519 words) Last reviewed: April 30, 2026 · Source · View sources On this page 12 sections Why DeepSeek’s papers matter more than most The DeepSeek paper timeline at a glance The foundation papers (2024) DeepSeek LLM — the scaling-law study DeepSeekMoE and DeepSeek-V2 — the architecture base DeepSeek-Coder and DeepSeekMath DeepSeek V3 — the efficiency thesis in paper form DeepSeek R1 — reasoning without human demonstrations DeepSeek V3.2 — the sparse attention bridge DeepSeek V4 — the current paper Architecture contributions Positioning against closed frontier models How the papers connect: a suggested reading order Using the papers as a developer, not a reader What the papers do not tell you Frequently asked questions Where can I find the official DeepSeek research papers? What does the DeepSeek V4 paper actually introduce? How much did it cost to train DeepSeek V3 and R1? Is the DeepSeek R1 paper really a reasoning trace breakthrough? Can I reproduce the benchmark numbers in the papers using the hosted API? Sources & Methodology Primary sources Technical references Context sources Methodology Data confidence Editorial note A Guide to DeepSeek Research Papers From V2 to V4 News · April 24, 2026 · By DS Guide Editorial If you are trying to understand why DeepSeek V4 uses only 10% of the KV cache of V3.2, or how a Chinese lab trained a reasoning model for under $300,000, the answer sits inside the DeepSeek research papers — not the press coverage. This guide maps every major publication the lab has put on arXiv and Hugging Face, from the original DeepSeek LLM scaling-law work in January 2024 to the DeepSeek-V4 technical report released on April 24, 2026. For each paper we note the model it describes, the single architectural idea it contributed, and how those ideas stack into the V4 release. You will finish with a reading order, a benchmar
|
||
=== https://hub.baai.ac.cn/view/54241 (7464 chars, 1 arxiv refs)
|
||
id: 2512.24880
|
||
TEXT: DeepSeek V4报告太详尽了！484天换代之路全公开 - 智源社区 活动 论文 知识树 风云榜 专栏 项目 社交 登录/注册 DeepSeek V4报告太详尽了！484天换代之路全公开 DDPM Gen AI DL 量子位 2026-04-26 09:40 分享 以下文章来源于mp.weixin.qq.com henry 发自 凹非寺 量子位 | 公众号 QbitAI DeepSeek V4“迟到”半年，但发布后的好评如潮还在如潮。 中外热搜上了一整圈，科技媒体的版面今天都让给了它，OpenAI也成了它的陪衬。 大家惊叹于DeepSeek在有限条件下作出重大突破的创造力，也佩服其在2026年，还能坚定选择开源路线的决心。 可以说，一时间信息多的有些超载，但多归多，主线就两条。 第一条， 百万token上下文全面开源，KV cache大幅缩减。 V4-Pro和V4-Flash， 1.6万亿参数/2840亿参数 ， 上下文都是1M 。1M场景下，V4-Pro的单token FLOPs只有V3.2的27%，KV cache只有10%。 亚马逊硬件师GPD表示，这意味着DeepSeek可能解决当前的HBM短缺问题。 第二条， 国产芯片适配 ，已经支持华为算力，预计下半年昇腾950超节点批量上市。 此外，大家最关心的，还莫过于在过去四个月中，DeepSeek陆续放出了几篇「可能进V4」的论文，今天技术报告开源了，可以对一下账。 mHC（流形约束超连接） ：2025年12月31日上传arXiv，梁文锋挂名。进了V4。 Engram（条件记忆模块） ：1月DeepSeek联合北大发布。没进V4，但在未来方向里被点名，留给V5。 DualPipe ：V3老伙计。继续用，针对mHC做了调整。 Muon优化器 ：从Kimi那边借的。V4把AdamW替了，接管绝大多数参数的训练。 四个预期，三个落地，一个给下一代。 整体架构 V4这一代，是DeepSeek系列里动刀最多的一版。相比V3，V4在三个地方做了升级。 第一，引入mHC（Manifold-Constrained Hyper-Connections）强化残差连接。 第二，设计hybrid attention架构，CSA和HCA交替叠加，解决长文效率问题。 第三，采用Muon作为主优化器。 MoE部分仍然用DeepSeekMoE，MTP（Multi-Token Prediction）模块跟V3保持一致。 一些细节微调包括，affinity score的激活函数从Sigmoid换成了Sqrt(Softplus(·))，去掉了routing target nodes的数量约束，前几层dense FFN换成了用Hash routing的MoE层。 接下来，我们一个个看。 mHC，给残差连接加一层约束 残差连接是 何恺明 2016年在ResNet里提出来的，十年没怎么变过。模型一层一层堆，梯度沿着残差往回传，这是深度学习能work的前提。 但模型越来越深、参数越来越多之后，传统残差开始露怯，信号传递不稳，训练容易崩。 先说Hyper-Connections（HC） ，这是Kimi团队之前提出的想法。核心是把残差流从一维变成n_hc条并行通道，每层之间通过一个矩阵B来混合。 A、B、C是三个线性映射。想法很优雅，相当于给残差流增加了一个新的scaling维度。但DeepSeek在堆多层时发现，HC经常出现数值不稳定，训练说崩就崩。 V4的做法叫mHC ，把矩阵B约束到「双随机矩阵」的流形上（数学上叫Birkhoff polytope），行和列都归一化为1。这个约束带来两个好处。 矩阵的谱范数天然不超过1，残差传播套上硬上限，爆不起来。 这种矩阵在乘法下是封闭的，堆很多层也稳。 输入映射A和输出映射C则通过Sigmoid函数保证非负且有界，避免信号互相抵消。 实现上用Sinkhorn-Knopp迭代，交替做行归一化和列归一化，迭代20次收敛。整个过程对每一层都跑一遍。 听起来贵，但DeepSeek做了fused kernel，再配合选择性recomputation，实测mHC带来的wall-time开销控制在overlapped pipeline的6.7%。 技术判断上，mHC不是那种让人眼前一亮的架构创新，更像是一个「稳得住大模型」的工程补丁。但随着模型深度和参数量继续往上推，这种补丁会变成刚需。 混合注意力机制 这是全篇论文最厚的一块，也是「百万token效率」的核心魔法所在。 V4的注意力层不是一种，是两种交替使用的结构，CSA（Compressed Sparse Attention）和HCA（Heavily Compressed Attention）。 CSA做两件事， 先压缩，再稀疏选择 。 第一步，KV压缩。 每m个token的KV entries，通过一个带学习权重的attention-like机制压成一个。 第二步，lightning indexer + top-k选择。 这部分继承自V3.2的DSA。对每个query token，用一个轻量的indexer计算它和每个压缩KV块的相关性分数。 第三步，core attention。 在选中的这top-k压缩KV块上做Multi-Query Attention，得到注意力输出。 第四步，grouped output projection。 因为V4把head dimension c设成了512（比V3.2的128大得多），如果直接把所有head的输出投影回d维会很贵，所以做了分组投影，把n_h个head分成g组，每组先投影到一个中间维度d_g，最后再合并投影回d。 整个CSA等于做了两层压缩。第一层是序列长度压缩，n变成n/m。第二层是稀疏选择，n/m变成top-k。对1M token的序列，原本需要attend 1M个token，现在只需要attend 1024个压缩块。 HCA的思路更简单粗暴，压得更狠，但不做
|
||
=== https://www.deepseek.com/ (524 chars, 0 arxiv refs)
|
||
TEXT: DeepSeek | 深度求索 获取 APP 扫码下载 DeepSeek App 中文 EN DeepSeek 网页版 DeepSeek 手机应用 API 开放平台 DeepSeek Harness API 文档 加入我们 中文 EN DeepSeek-V4.1-Flash 发布，文本与 Agent 性能全面提升、兼具原生多模态视觉理解能力，能力更强、速度更快、且成本更低，欢迎测试和反馈 探索未至之境 深度思考 智能搜索 和 DeepSeek 对话 使用 API 开放平台 浙ICP备2023025841号 浙B2-20250178 浙公网安备33010502011812号 扫码关注公众号 产品 DeepSeek 网页版 DeepSeek 手机应用 DeepSeek Harness API 开放平台 API 文档 API 价格 服务状态 研究 DeepSeek V4.1 DeepSeek V4 DeepSeek V3.2 DeepSeek V3.1 DeepSeek R1 DeepSeek V3 更多 法务 & 安全 反馈安全漏洞 透明度 加入我们 岗位详情 © 2026 杭州深度求索人工智能基础技术研究有限公司 版权所有 中文 EN |