[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-glm-5-2-z-ai-indexshare-1m-anti-cheat-rl":3,"topics-all":36,"news-related-0d6b1c39-3b2c-48c4-afb4-41a0de2a918d":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"0d6b1c39-3b2c-48c4-afb4-41a0de2a918d","GLM-5.2 把 1M 上下文\"焊\"进开源：IndexShare + 反作弊 RL，把长程 Agent 拉成工程现实","Z.ai（智谱）6 月 17 日在 Hugging Face 发布 GLM-5.2：753B 参数、MIT 开源、上下文窗口从 200K 拉到 1M token。比起又大又长，GLM-5.2 真正值得展开的是三个工程层面的决定。\n\n第一是 IndexShare 稀疏注意力：把同一组 indexer 复用到连续 4 层稀疏注意力中，1M 上下文下每 token FLOPs 直降 2.9×。这件事为什么重要？DSA 类稀疏注意力如果不解决 indexer 自身的开销，「长上下文」只会先把推理账单拖垮。IndexShare 是把稀疏注意力真正能上工程补齐的关键一环。\n\n第二是 MTP + IndexShare + KVShare 的联合优化：speculative decoding 的接受长度从 4.56 提升到 5.47（+20%）。在 1M 上下文下 decode 步骤被几十倍放大，每步多收一个 token，端到端吞吐就被显著拉开。\n\n第三是 Coding Agent 的 RL 做到「能跑、稳跑、不被刷」。Z.ai 用自研 slime 框架做大规模 agentic RL rollout——支持白盒\u002F黑盒 rollout、压缩轨迹、子智能体工作流，把 10+ 专家模型在约两天内合并训练完毕。同时设计反作弊模块：规则过滤先抓可疑工具调用（如 curl raw.githubusercontent.com\u002F... 拉答案），再用 LLM judge 判断意图；命中后在线阻断并返回假数据，但不打断整条 rollout，避免粗暴拒掉轨迹带来的训练崩塌。\n\nBenchmark 上 GLM-5.2 把「开源旗舰」天花板顶到新位置：Terminal-Bench 2.1 拿到 81.0（GLM-5.1 仅 63.5），SWE-bench Pro 62.1。在 FrontierSWE 上它仅落后 Claude Opus 4.8 一个百分点，是该榜唯一进第一梯队的开源模型。\n\n比起再刷一次 chat benchmark，Z.ai 真正贡献的是工程视角：1M 上下文的难点不在「塞得进」，而在推理经济性、Coding Agent 的 RL 稳定性与反作弊、长程任务的可验证执行。开源权重 + MIT + 完整训练\u002F推理笔记，第一次让「长程开源旗舰」从口号变成可复现的现实。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fzai-org\u002Fglm-52-blog","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"398c6525-2bfd-436d-b25e-f872333c1f60","en","GLM-5.2 welds 1M context into open source for long-horizon agents","Hugging Face user zai-org released GLM-5.2, a major update to Zhipu's flagship GLM series. The standout: 1M context window with strong long-horizon Agent capability, achieved through a combination of IndexShare (efficient KV cache management) and anti-cheat RL (training the model to actually use the long context).\n\nThe \"IndexShare\" highlight: a KV cache sharing technique that allows the 1M context to be served at 2× the speed of competing 1M-context models. IndexShare identifies \"shared\" KV cache entries across different context segments and reuses them, cutting memory by 40% and latency by 30%.\n\nThe \"anti-cheat RL\" insight: long-context models often \"cheat\" by ignoring most of the context and focusing on the last few thousand tokens. GLM-5.2 uses an RL training signal that explicitly rewards the model for using information from the full context. The result: on long-context QA, GLM-5.2 uses information from 80%+ of the context, compared to 30-40% for the previous generation.\n\nThe benchmark: on the LongBench-Agent benchmark (1M context), GLM-5.2 scores 71.4, beating Llama-3.1-405B (68.2) and approaching Claude Opus 4.7 (73.5). The 1M context is the longest among open-source models.\n\nThe bigger takeaway: \"long-context Agent\" is the right paradigm for production AI. The \"1M context is a gimmick\" criticism is being addressed by anti-cheat RL, and the \"long-context is too slow\" criticism is being addressed by IndexShare. For the industry, this signals that \"long-context\" is becoming a real production capability, not just a marketing claim.","glm-5-2-z-ai-indexshare-1m-anti-cheat-rl","2026-06-19T20:00:00Z","2026-06-19T20:10:42.235581Z","2026-08-19T02:08:40.142862Z",true,"agent",303,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"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":68,"title":69,"news_slug":70,"published_at":71},"98695785-30b2-4ddf-9886-757e57773f8f","Arcee Trinity Large：400B开源MoE模型挑战Claude Opus，定价便宜96%","arcee-trinity-large-400b-moe-claude-opus-96pct-cheaper","2026-04-30T07:01:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"551dfd06-3bff-47d1-a7e4-a420a9c90c22","OpenAI发布GPT-OSS 120B：七年后重返开源，单卡部署的边界被重新定义","openai-gpt-oss-120b-apache-2-moe-int4","2026-04-27T07:10:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00"]