[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-zhipu-glm-rsi-infrastructure-chinese-cluster":3,"topics-all":38,"news-related-0fe869ca-11c4-4831-afb1-19a31fd88dfc":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"0fe869ca-11c4-4831-afb1-19a31fd88dfc","智谱公开国内大模型首个 RSI:GLM-5.3 Infra Agent 在 10 万国产卡集群自建推理,2 周吞吐 3 倍","智谱 9 月 17 日公开国内大模型首个 RSI 实践:由 GLM-5.3 驱动的 Infra Agent 在 10 万张国产加速器上从零搭建生产级推理服务,2 周内端到端吞吐提升 3 倍,硬件利用率与单 token 成本对齐主流 NVIDIA GPU。","智谱今天在官方技术博客公开了国内大模型厂商里的第一个 RSI（Recursive Self-Improvement）工程实践：他们让 GLM-5.3 驱动的 Infra Agent，从零搭建并优化了 GLM-5.3-Flash 的整套生产级推理服务。生产环境跑在 10 万张以上国产 AI 加速器上，不到两周端到端吞吐就做到了初始基线的 3 倍，硬件利用效率和单 token 成本对齐了主流 NVIDIA GPU 的水平。\n\nRSI 的字面意思是「递归自我改进」——模型不断优化自己运行的系统，再由这套系统去支撑下一代模型的训练，听起来离现实很远。但智谱这次给出的案例说明：早期形态已经在生产环境里跑通了。\n\n要把一个新模型从「在国产硬件上第一次跑通」推到「能扛生产流量的高吞吐服务」，历来是几周甚至几个月的系统工程。智谱这次的难点在于三件事叠加：芯片显存与带宽受限、要支持 1M token 上下文和图文多模态、生态不成熟、kernel 支持残缺。能跑出这个成绩，是因为 Infra Agent 承担了过去由基础设施工程师手工完成的大部分工作。\n\n## 关键机制：把工程师经验拆成「可归因的反馈」\n\n团队把核心创新归结为一套他们称之为「dense feedback」的反馈机制，不是堆日志、堆 metric，而是让每一条反馈都能归因到一个具体的引擎参数、代码改动、kernel、输入条件或代码路径。三类反馈对应三类决策：正确性反馈告诉智能体「算的对不对」，系统行为反馈定位「时间花在哪」，性能反馈回答「哪种方案更好」。这三层组合起来，把过去依赖资深工程师经验串联起来的诊断过程，变成智能体可以连续执行的工程闭环。\n\n博客给了三个具体案例。\n\n## 案例一：把 KDA kernel 的累积误差修回容差内\n\n智能体通过对比并行与非并行 kernel 的输出，发现 KDA kernel 的 Context Parallelism 路径在状态合并时累积误差。根因是 tl.dot 默认走 TF32，把输入精度从 tf32 改成 tf32x3 之后误差被压回容差内。这条修复已经合并到 Flash Linear Attention 的 PR #1180。\n\n## 案例二：补一个 GIL 释放，KV Transfer 差距从 20% 压到 1%\n\nPrefill + KV Transfer 的吞吐和 Prefill-only 的差距超过 20%，远超 5% 的目标。智能体沿着时间线追到 DeepEP v1.2.1 的两个 C++ 函数没有显式释放 GIL，把同一进程里的 Mooncake Transfer 提交线程堵死。补上 GIL 释放后差距压到 1% 以下。\n\n## 案例三：合并沿 V 维的冗余 tile，单 kernel 提速 1.71 倍\n\n智能体从 SGLang、Flash Linear Attention、DeepGEMM 等开源 kernel 里提炼出「优化骨架」，再结合 profiling 重新评估 tiling、内存访问与资源分配。在 KDA Decode kernel 上，靠合并沿 V 维的冗余 tile，单 kernel 拿到 1.71 倍提速。\n\n## 真正的信号\n\n智谱在博客里承认：选目标、设边界、做风险评估仍然是人。但 2 周、3 倍吞吐、10 万张国产卡，这三个数字说明一件事——模型优化系统、系统跑模型，这条循环一旦在生产里跑顺，前沿模型迭代对底层工程团队规模的依赖会迅速下降。这才是 RSI 真正值得关注的信号。\n","https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-built-its-inference-infrastructure","df9ef325-77c5-4e95-9c03-f6cf5b150ef0",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"494ef097-5bd7-454e-b426-2a1c83f10e67","en","Zhipu unveils first China LLM RSI run on 100K Chinese chips","GLM-5.3 Infra Agent built production inference on 100,000+ Chinese AI accelerators, hitting 3x throughput in two weeks at NVIDIA parity.","Zhipu today published the first engineering RSI (Recursive Self-Improvement) case from a Chinese LLM vendor on its official technical blog. A GLM-5.3-powered Infra Agent designed, debugged, and optimized the entire production-grade inference stack for GLM-5.3-Flash from scratch, deployed on a cluster of more than 100,000 Chinese-made AI accelerators. End-to-end throughput tripled relative to the initial baseline in under two weeks, while hardware utilization and per-token cost reached levels comparable to mainstream NVIDIA GPUs.\n\nRSI in name refers to systems that iteratively improve themselves and then help train their own successors. That endpoint still looks far off, but the early form is now running in production.\n\nGetting a new model from its first successful run on unfamiliar hardware to a high-throughput inference service that can reliably handle production traffic has historically taken a team of infrastructure engineers weeks or months. Zhipu stacked three complications this time: limited chip memory capacity and bandwidth, a 1M-token context window, multimodal traffic, an immature ecosystem, and incomplete kernel support. The Infra Agent carried most of the work that used to live with the systems team.\n\n## The core idea: dense feedback that engineers can still sign off on\n\nThe team's central contribution is what they call \"dense feedback,\" a feedback discipline that ties every signal back to a specific engine launch parameter, code change, kernel, input condition, or code path, rather than dumping more logs and metrics onto the agent. Three kinds of feedback answer three kinds of decisions: correctness feedback tells the agent whether the computation is right; system-behavior feedback pinpoints where time is going; performance feedback decides which approach actually wins under the target workload. The three layers convert the diagnostic chain that used to live inside an experienced engineer's head into a workflow an agent can execute continuously.\n\nThe blog walks through three concrete cases.\n\n## Case 1: numerical accuracy in a KDA kernel\n\nThe agent compared partitioned vs. unpartitioned kernel outputs and uncovered an accumulating-error bug on the KDA kernel's Context Parallelism path. The root cause was tl.dot defaulting to TF32 even with FP32 inputs, which compounded during state merging and updates. Setting input_precision=\"tf32x3\" pulled the error back within tolerance. The fix was merged upstream into Flash Linear Attention as PR #1180.\n\n## Case 2: a missing GIL release that blocked KV Transfer\n\nThe performance gap between Prefill + KV Transfer and Prefill alone exceeded 20%, well past the 5% acceptance bar. Tracing the execution timeline showed that the Python Mooncake Transfer thread in the same process could not acquire the GIL promptly because DeepEP v1.2.1's intranode_dispatch and intranode_combine never released it explicitly. Releasing the GIL during the relevant C++ intervals closed the gap to under 1%.\n\n## Case 3: collapsing redundant V-dim tiles for a 1.71x kernel speedup\n\nThe agent extracted optimization techniques from existing kernels across SGLang, Flash Linear Attention, and DeepGEMM, then turned them into \"optimization skeletons\" with applicability conditions, transformation methods, and validation evidence. On a representative KDA Decode kernel, after ReplaySSM and a 9.6% reduction from division optimization, the agent noticed the original implementation tiled along the V dimension, repeating FP32 normalization and gating four times. Merging those tiles into a single thread block and replacing per-tile reductions with a single warp-level reduction yielded a 1.71x speedup over the v2 baseline.\n\n## The real signal\n\nZhipu is clear that choosing objectives, setting boundaries, and assessing production risk remain human responsibilities. But two weeks, threefold throughput, and 100,000 accelerators together say one thing: once the loop of \"model optimizes system, system runs model\" works in production, the engineering headcount required to ship the next frontier model starts to drop fast. That is the RSI story actually worth watching.\n","zhipu-glm-rsi-infrastructure-chinese-cluster","2026-09-17T08:00:00Z","2026-09-20T03:12:47.032074Z","2026-09-20T03:12:47.032092Z",true,"agent",35,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b0183d10-bcfd-44ed-a178-a2c813f10b69","国家超算互联网AI社区上线Kimi K3:2.8万亿参数MoE一键调用,开源大模型有了国产算力底座","kimi-k3-scnet-platform-launch","2026-07-28T09:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d9f47040-7c31-4739-8d3e-23acc06162d2","爱芯元智 Pulsar2 6.0 + axllm：把端侧 LLM 工具链拼成云上同款","axera-pulsar2-6-0","2026-06-30T08:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06","Mistral 模型进驻 Firefox:119B 开放权重模型驱动浏览器 AI 助手","mistral-small-4-firefox-smart-window","2026-09-16T17:07:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b7668f43-05a7-46d7-845f-27e70fcaceec","Mozilla 报告:中国开放权重距美国前沿模型只差 4.4 个月","mozilla-open-weights-4-month-gap","2026-09-16T13:08:54+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"e84fe968-5d86-4247-baad-5da23efef860","UltraData-RL-2609 开源:85,995 条可验证奖励任务,拆解 MiniCPM5-2B 的 RL 燃料","ultradata-rl-2609-verifiable-rl-dataset","2026-09-07T23:07:45+00:00"]