[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v41-flash-kv-cache-paper":3,"topics-all":41,"news-related-b667e52f-ec7d-4ca4-8d9e-1db81e1a5616":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","DeepSeek 发布 V4.1-Flash 技术报告:552B MoE 把每 token 全局 KV 缓存压到 890 字节,CED 让 prefill 只激活 8B,CSA2 加 FP4 砍到上代 1\u002F4,SWA Bounded Replay 把持久缓存再压到 1\u002F8。Agent 基准官方口径领先,MIT 开源。","长程 agent 的工作负载有一个共同特征:输入远多于输出。一份代码库、一沓文档、几十轮工具调用记录塞进上下文,prefill 的算力、HBM 里的 KV cache、SSD 的容量和搬运带宽,三样同时吃紧。DeepSeek 9 月 17 日放出的 V4.1-Flash 技术报告([arXiv:2609.19969](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19969)),通篇就在算一笔账:同一个 552B 的模型,怎么把每个 token 的 KV 缓存开销压到 890 字节。\n\n## 第一层:CED 非对称激活,读题用 8B、答题用 16B\n\nV4.1-Flash 采用 Causal Encoder-Decoder(CED)架构:40 层 Transformer,前 20 层做因果编码器,后 20 层做解码器。关键在于,解码器的全局 KV cache 不再由各解码层自己的隐藏状态推导,而是直接从编码器最终的隐藏状态投影而来。好处立竿见影:prefill 阶段每个 token 只需激活 8B 参数,decode 阶段激活 16B。对输入重、输出轻的 agent 负载,省的正是最贵的那一段。9 月 10 日 API 公告里「读题 8B 答题 16B」的说法,在这份报告里给出了完整的架构依据。\n\n## 第二层:CSA2 三模式加 FP4,全局缓存砍到上代 1\u002F4\n\nCompressed Sparse Attention 2(CSA2)给每个注意力层分配三种静态模式之一——Full、Reindex 或 Reuse——让多层共享主 KV 与索引 K,并复用 Top-K 稀疏注意力索引。解码器里的分层稀疏索引器(Hierarchical Sparse Indexer)进一步把后续索引层的候选池限制在第一个 Full 模式层构造的集合内,更深层索引器的成本不再随上下文长度增长。再叠上 FP4 主 KV 缓存(E2M1 格式,每 16 个通道共享一个 E4M3 缩放因子),全局 KV 缓存落到每 token 890 字节——约为 V4-Flash 的 1\u002F4、V1 的 1\u002F437。\n\n## 第三层:SWA Bounded Replay,持久缓存再砍到 1\u002F8\n\n驻留在 SSD 或主机内存里的持久 KV 是另一笔开销。SWA Bounded Replay 的思路是:不把滑动窗口注意力的 KV 状态持久化,需要时只重放最近 n_win 个 token 重建缺失状态。持久缓存占压到 V4-Flash 的约 1\u002F8。\n\n## 架构外的配套件\n\n报告还交代了几个值得注意的组件:Engram 条件记忆模块(196B 参数,按 token 查找的稀疏访问);DSpark 投机解码(半自回归草稿生成,置信度调度验证);MoE 每层 1 个共享专家加 384 个路由专家、每 token 激活 6 个。预训练从零开始,45T token 多模态语料,稀疏注意力在 64K 序列长度上训练,34T 处扩到 1M 上下文。后训练沿用 SFT、RL 到在线蒸馏(OPD)的流程,改动集中在数据管线——agent 任务与环境的大规模自动合成。推理侧支持 1-100 整数档的连续可控推理强度。\n\n## 官方口径的跑分:agent 场景领先,硬解推理仍有差距\n\n按模型卡公布的基准(官方内部框架评测口径),最大推理强度下:Terminal-Bench 2.1 90.6(高于 Opus-5.0 的 89.1)、DeepSWE v1.1 74.2(高于 Opus 74.0)、CyberGym 88.1、Agent's Last Exam 31.8、AutomationBench 54.8,均为对比表中最优;Codeforces 评分 3471 也是全表最高。但差距同样写在表里:Terminal-Bench 3.0\u002F4.0 分别只有 30.0\u002F31.2,明显落后 Opus 的 43.3\u002F51.8;HLE 36.8 对 Opus 的 56.3;GPQA 90.9 低于 GPT-5.6 Sol 的 94.1。便宜不等于全能,这份把短板一并放出来的表格,比单看几项第一更有信息量。\n\n## 工件配套:这次连 Rust 库都给了\n\n[HF 仓库](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-V4.1-Flash)(MIT 许可)除权重外,还带了 prompt 编码参考实现、推理代码、DeepSWE 复现步骤,以及 deepseek-recipe——一组带 Python 绑定的 Rust 库,统一处理 V4 与 V4.1 的 prompt 编解码。社区侧已有 73 个量化版本、14 个微调模型跟进。值得注意的是这次没有提供 Jinja 格式 chat template,改用自维护的编码工具链——新架构的 prompt 格式复杂度,可能已经到了通用模板不好覆盖的程度。\n\n三层缓存设计不是把模型缩小,而是把「上下文怎么存」重新设计了一遍。当 agent 把百万 token 上下文当常规配置用,HBM 和 SSD 的账本迟早会摆到每家推理团队桌上。这份报告给出的参照系是:架构层的非对称激活、注意力层的跨层共享、部署层的有损重建,三件事可以叠加,而且开源可验证。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19969","4194681c-1a38-405d-a917-40e1dc2622ea",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":25,"name":26,"slug":26,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"3b75b975-f7d6-4e6c-b3cf-c78baa36540c","en","DeepSeek's 890-Byte KV Cache Paper, Decoded","DeepSeek's V4.1-Flash technical report details how a 552B MoE cuts per-token global KV cache to 890 bytes: CED activates only 8B params in prefill, CSA2 plus FP4 trims cache to 1\u002F4 of the previous generation, and SWA Bounded Replay cuts persistent cache to 1\u002F8. Weights are MIT-licensed.","Long-horizon agents share one workload trait: input dwarfs output. A codebase, a stack of documents, dozens of tool-call transcripts crammed into context — prefill compute, KV cache in HBM, SSD capacity and transfer bandwidth all strain at once. DeepSeek's V4.1-Flash technical report, released September 17 ([arXiv:2609.19969](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19969)), is essentially one accounting exercise: how does a 552B model push per-token KV cache overhead down to 890 bytes?\n\n## Layer 1: CED asymmetric activation — 8B to read, 16B to answer\n\nV4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. The key move: the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. The payoff is immediate — only 8B parameters activate per token during prefill, 16B during decode. For input-heavy, output-light agentic workloads, the savings land exactly on the most expensive segment. The \"8B to read, 16B to answer\" framing from the September 10 API announcement now has its full architectural justification.\n\n## Layer 2: CSA2's three modes plus FP4 — global cache cut to 1\u002F4 of the previous generation\n\nCompressed Sparse Attention 2 (CSA2) assigns each attention layer one of three static modes — Full, Reindex, or Reuse — letting layers share main KV and indexer K while reusing Top-K sparse-attention indices. A Hierarchical Sparse Indexer in the decoder further restricts later indexing layers to a candidate pool built by the first Full-mode layer, bounding deeper indexer cost independently of context length. Layered on top, FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels) brings the global KV cache footprint to 890 bytes per token — roughly 1\u002F4 of DeepSeek-V4-Flash and 1\u002F437 of V1.\n\n## Layer 3: SWA Bounded Replay — persistent cache cut to 1\u002F8\n\nPersistent KV residing on SSD or in host memory is a separate cost. SWA Bounded Replay skips persisting sliding-window-attention KV states entirely; when needed, it reconstructs missing states by replaying only the most recent n_win tokens. The persistent cache footprint drops to roughly 1\u002F8 of V4-Flash.\n\n## Components beyond the cache stack\n\nThe report also details several notable components: Engram conditional memory (196B parameters, sparsely accessed via token-based lookup); DSpark speculative decoding (semi-autoregressive draft generation with confidence-scheduled verification); and an MoE design of 1 shared expert plus 384 routed experts per layer, activating 6 routed experts per token. Pretraining ran from scratch on a 45T-token multimodal corpus, with sparse attention trained at 64K sequence length and context extended to 1M tokens at the 34T mark. Post-training follows the standard SFT → RL → on-policy distillation (OPD) recipe with no algorithmic changes — the substantive work moved to the data pipeline: large-scale automated synthesis of agent tasks and environments. Inference supports a continuously controllable reasoning-effort setting, an integer scale of 1 to 100.\n\n## Officially reported benchmarks: leading on agents, still behind on hard reasoning\n\nPer the model card's benchmarks (evaluated in DeepSeek's internal framework, vendor-reported), at maximum reasoning effort: Terminal-Bench 2.1 at 90.6 (above Opus-5.0's 89.1), DeepSWE v1.1 at 74.2 (above Opus's 74.0), CyberGym 88.1, Agent's Last Exam 31.8, AutomationBench 54.8 — all best in the comparison table; the Codeforces rating of 3471 is also the table's highest. But the gaps are right there in the same table: Terminal-Bench 3.0\u002F4.0 at 30.0\u002F31.2 trail Opus's 43.3\u002F51.8 by a wide margin; HLE 36.8 versus Opus's 56.3; GPQA 90.9 below GPT-5.6 Sol's 94.1. Cheap is not omniscient — a table that publishes its own weak spots alongside the wins carries more information than any single first-place finish.\n\n## Artifacts included: even the Rust libraries shipped\n\nThe [HF repository](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-V4.1-Flash) (MIT license) ships weights plus a prompt-encoding reference implementation, inference code, DeepSWE reproduction steps, and deepseek-recipe — a set of Rust libraries with Python bindings that uniformly handles prompt encoding and decoding for V4 and V4.1. The community has already followed with 73 quantized variants and 14 fine-tunes. Notably, no Jinja chat template is provided this time; the encoding toolchain is self-maintained instead — a hint that the new architecture's prompt-format complexity may have outgrown generic templating.\n\nThe three-layer cache design doesn't shrink the model; it redesigns how context is stored. As agents treat million-token contexts as routine configuration, the HBM-and-SSD ledger will land on every inference team's desk sooner or later. The reference frame this report offers: asymmetric activation at the architecture layer, cross-layer sharing at the attention layer, and lossy reconstruction at the deployment layer — three moves that stack, and that are open to verify.","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00Z","2026-09-18T15:11:47.559945Z","2026-09-18T15:11:47.559961Z",true,"agent",202,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"823e17ef-5927-40e6-9efd-08c4958922f0","DeepSeek 旗舰 V4-Pro 今日退役:552B 的 V4.1-Flash 全面接班","deepseek-v4-pro-retires-flash-routes","2026-09-14T15:10:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"ee62535f-b897-437b-8674-02801632dadb","DeepSeek V4.1 非对称架构首发:读题 8B 答题 16B,KV 缓存砍到初代的 1\u002F437","deepseek-v4-1-flash-ced-kv-cache","2026-09-10T15:10:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","minimax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"b571067a-9fa8-42bf-9431-98f26ac78e03","伯克利把LLM推理搬进SSD:KV缓存压缩15倍","llm-inference-in-flash-cim-ssd","2026-09-19T21:10:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"8c5bd0e1-551a-405b-b2fc-a67520bf84c5","Colibri v1.11.0 发布:纯 C 引擎直读 DeepSeek V4.1 Flash,552B 从 SSD 流进 CPU","colibri-v1-11-deepseek-v41-flash","2026-09-14T13:09:40+00:00"]