[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-glm-5-3-post-training-emergent-cyber":3,"topics-all":38,"news-related-9f800588-ad3b-4eee-a8a6-e2db1ac8f014":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},"9f800588-ad3b-4eee-a8a6-e2db1ac8f014","GLM-5.3:只靠后训练把 743B 基座打成新 SOTA,网络安全的\"涌现\"打了 Z.ai 一个措手不及","8 月 14 日,Z.ai 发布 GLM-5.3。基座与 GLM-5.2 完全相同(743B),所有能力提升都来自\"更大规模后训练\":更多任务环境、更多训练算力。Terminal-Bench 3.0 从 4.6 飙到 28.3,DeepSWE v1.1 从 46.2 涨到 66.9。更让团队意外的是,CyberGym 上从 77.2% 涨到 84.5%(超过 Mythos 5 的 83.8% 和 GPT-5.6 Sol 的 83.6%),且越深入漏洞利用链,涨幅越大——能力恰好增长得最快的地方,正是离闭源前沿差距最大的地方。权重两周后公开,slime RL 框架已开源。","# GLM-5.3:只靠后训练把 743B 基座打成新 SOTA,网络安全的\"涌现\"打了 Z.ai 一个措手不及\n\n**一句话摘要**:Z.ai 用同一份 743B 基座、只升级后训练栈,把 GLM-5.2 推进到 GLM-5.3——长程编码与网络安全两个维度同时跨入 SOTA 区;但这次最大的新闻,是 Z.ai 自己都没预料到的\"涌现\"。\n\n---\n\n## 背景:从 5.2 到 5.3,基座一字未动\n\n2026 年 8 月 14 日,Z.ai(智谱的海外品牌)正式推出 GLM-5.3。最大的反常识点是:基座模型与 GLM-5.2 完全相同——同样 743B 参数,同样架构,同样权重。Z.ai 在官方博客里直接说,\"Every gain comes from post-training\"。\n\n支撑这次升级的,是 GLM-5.2 时代搭起来的三块基础设施:\n\n- **IndexShare**:处理长上下文的高效注意力机制\n- **SAO**(Stale-Aware Optimization):用于长程任务的强化学习算法\n- **slime**:Z.ai 开源的大规模异步 RL 训练框架(训练端跑 Megatron,推理端跑 SGLang)\n\n过去一个月,团队就在这块底座上做加法:更多任务环境、更多环境类型、更长训练时长。从结果看,这种\"环境规模换能力\"的路线非常有效。\n\n## 编码维度:长程任务涨幅最大\n\nGLM-5.3 在编码类基准上的跃升呈现一个清晰特征:**任务越长,涨得越凶**。\n\n几个关键数字:\n\n- **Terminal-Bench 3.0**:4.6 → 28.3(超过 6 倍)\n- **DeepSWE v1.1**:46.2 → 66.9\n- **Agents' Last Exam(CLI 版)**:23.8 → 28.5\n- **FrontierSWE**:67.5 → 78.1\n- **GDPval-AA v2**(覆盖 44 个职业):1508 → 1769\n\n在自家 Z.ai Code Bench 上,GLM-5.3 相对 GLM-5.2 提升 50%。Max effort 下,以约 75K 输出 token 拿到 34.5% 完成率,而 GLM-5.2 要 96K token 才打到 23.4%。High effort 下,以约 50K token 拿到 31.4%,超过了 Claude Opus 4.8 用 120K token 拿到的 29.5%——**用更少 token,跑赢闭源旗舰**。\n\n当然,Claude Fable 5 在 Max effort 下依然领先(39.5%),而 GPT-5.6 Sol 在 Terminal-Bench 3.0 上仍以 34.6% 略压 GLM-5.3 的 28.3%。\n\n## 网络安全:\"涌现\"打了 Z.ai 自己一个措手不及\n\n这才是 GLM-5.3 真正让行业意外的部分。\n\nZ.ai 在博客里直白承认:原本只是想\"让模型找漏洞更准\"。他们往训练 mix 里加了漏洞发现数据和对应环境,期望看到的是\"更会找漏洞\"。**结果是,能力不止于此**。\n\n> \"What surprised us was how quickly the capability continued to develop as training scaled.\" ——Z.ai 官方博客\n\n具体的越级表现:\n\n- **CyberGym**(白盒源码漏洞发现):77.2% → **84.5%**,超过 Mythos 5(83.8%)和 GPT-5.6 Sol(83.6%),位列榜首\n- **ExploitBench**(需要更深层推理的真实漏洞利用):24.4% → **54.4%**,翻倍以上\n- **ExploitGym**(在时间预算内完成的利用任务数):2 小时内 29 → **105**,6 小时内 39 → **130**\n- **Mythos 5 在 ExploitBench 上仍以 78.0% 领先,在 ExploitGym 上 2h\u002F6h 分别完成 181\u002F247 任务**\n\nZ.ai 给出的判断是:**任务离\"真实利用链\"越远,涨幅越大;越深、利用链越完整,涨幅越夸张**。换句话说,模型不只是\"找到了 bug\",而是开始**形成跨阶段的完整利用计划**。\n\n更值得玩味的是这种增长分布:**能力增长得最快的地方,恰好是 GLM-5.3 与闭源前沿差距最大的地方**。CyberGym 上 GLM-5.3 已经超车,但 ExploitBench 上还差 Mythos 5 约 24 个百分点,ExploitGym 上 Mythos 5 的完成数仍是其近两倍。\n\n## 真实世界验证:269 个项目、2436 个漏洞、最早可追溯到 1981 年\n\n为了检验这些能力是否过拟合在 benchmark 上,Z.ai 自 GLM-5.2 起就与中国若干安全团队合作,**让模型在真实代码库上找漏洞**。\n\n经过专家复核、筛选、去重后的统计:\n\n- 269 个开源项目\n- 2436 个漏洞发现\n- 其中 1097 个为中\u002F高危\n- 覆盖系统内核、操作系统、浏览器引擎、开源基础设施、Web 应用、网络协议\n- **最早的一个漏洞可追溯到 1981 年**,平均每个漏洞潜伏 26.6 年才被发现\n\nZ.ai 为此专门搭了 [Security Disclosure Ledger](https:\u002F\u002Fcvd.z.ai\u002F) 公开披露台账,实时跟踪披露进度。截至发稿,53 个已公开、2383 个仍处于披露前阶段。\n\n## slime 框架:把 2.3× 训练吞吐喂进去的\"水电煤\"\n\n上述所有结果都跑在 Z.ai 开源的 [slime](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime) 框架上。核心设计是训练、rollout、数据 buffer 跑在同一条 dataflow 上,让 RL 环境的接入变成\"数据生成\"而不是\"重写训练 loop\"。\n\n这次同步升级的两个维度:\n\n1. **算法侧**:加了 top-p mask、top-k 与 full-vocabulary OPD、以及改进训练-推理一致性的 R3-style 配置。训练-推理一致性评估里,平均 logprob 差异被压到了 **1e-7 量级**(相对之前降低 99.99% 以上)\n2. **系统侧**:本地存储成为额外缓存层、多教师 OPD 不再需要长跑推理服务、长程任务的端到端训练吞吐提升 **超过 2.3×**\n\n换句话说,GLM-5.3 的能力不是基座换来的,是\"环境 + 算法 + 系统\"三件事同步加码换来的。\n\n## 怎么用 + 一个\"坑\"\n\n- 已在 [Z.ai API](https:\u002F\u002Fdocs.z.ai\u002Fguides\u002Fllm\u002Fglm-5) 与 GLM Coding Plan 全面开放\n- 支持 Claude Code、OpenCode 等主流 coding agent harness\n- 权重两周后公开,需先完成安全评估与硬化\n- **API 改动**:GLM-5.3 不再支持 ,统一为  + 。迁移代码若没改这个,请求会直接失败\n- GLM Coding Plan 改成按点数计费,14:00–18:00(UTC+8)工作日为高峰时段,非高峰 token 消耗打 5 折\n\n## 个人评论\n\nGLM-5.3 给行业留下的最大问题不是\"GLM 5.3 有多强\",而是:\n\n1. **后训练规模化是否还有边际收益?** 同一个基座 743B 上,Gym 类基准 2 个月涨这么多——这意味着我们之前对\"基座能力决定上限\"的判断可能是错的,RL 环境规模和质量的曲线还很陡\n2. **涌现式网络安全的\"安全悖论\"**:模型越来越会找漏洞,这正是 Anthropic Mythos 5 严格限制发布的原因。Z.ai 选择\"先在线开放,2 周后再发权重\",并搭了公开披露台账——这条路线如果跑通,可能是开放模型在 cyber 维度上第一次把\"披露流程\"作为一等公民\n3. **Cyber 能力的代差分布**:Z.ai 在博客里自己点出的现象——\"能力增长最快的地方,就是差距最大的地方\"——意味着开源 vs 闭源在 cyber 上的差距不会自然收敛,只会按任务的\"实战深度\"重新分布\n\n短期看,这对做应用安全、做 CI 修复、做代码审计的团队是个明确信号:长程 Agent 已经在 benchmark 上接近闭源前沿,产品化窗口期可能是接下来这 2-3 个月。\n\n长期看,这次发布最大的受益者可能不是任何一家模型厂商,而是所有用了 GLM-5.2 的产品——后训练的可堆叠性,意味着我们已经可以从一份\"足够好的基座\"上,反复用 RL 环境把它推到新位置,而不必每次都烧一遍预训练。\n\n参考资料:\n- [GLM-5.3 官方技术博客](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3)\n- [MarkTechPost 报道](https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F08\u002F14\u002Fz-ai-ships-glm-5-3-without-retraining-the-base-model-better-at-complex-coding-and-long-horizon-tasks\u002F)\n- [slime 开源框架](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime)\n- [Z.ai Security Disclosure Ledger](https:\u002F\u002Fcvd.z.ai\u002F)","https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3","df9ef325-77c5-4e95-9c03-f6cf5b150ef0",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"44b5626a-85a7-40ae-a0ab-bc9c1433fe26","en","GLM-5.3: post-training turns a 743B base into the new SOTA","On August 14, 2026, Z.ai released GLM-5.3 using the same 743B base as GLM-5.2 — every capability gain came from larger-scale post-training: more task environments, more environment types, more compute. Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9. The bigger surprise came on cybersecurity: CyberGym climbed from 77.2% to 84.5% (ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%), and the deeper the exploitation chain, the larger the gain — capability is growing fastest exactly where the gap to closed-frontier models is widest. Weights public in two weeks; the slime RL framework is already open-source.","# GLM-5.3: Post-Training Alone Lifts a 743B Base to New SOTA, as Emergent Cyber Capability Surprised Z.ai\n\n**TL;DR**: Z.ai pushed GLM-5.2 to GLM-5.3 without changing the base model — every gain came from scaling post-training on the same 743B foundation. Long-horizon coding and cybersecurity both crossed into SOTA territory. But the real headline is the capability Z.ai did not plan for.\n\n---\n\n## Background: From 5.2 to 5.3, the Base Did Not Move\n\nOn August 14, 2026, Z.ai (the international brand of Zhipu) released GLM-5.3. The most counter-intuitive fact: the base model is byte-for-byte the same as GLM-5.2 — same 743B parameters, same architecture, same weights. Z.ai's official blog puts it bluntly: \"Every gain comes from post-training.\"\n\nThree pieces of infrastructure built during the GLM-5.2 era made this possible:\n\n- **IndexShare**: efficient attention for long-context processing\n- **SAO** (Stale-Aware Optimization): an RL algorithm designed for long-horizon tasks\n- **slime**: Z.ai's open-source asynchronous RL training framework (Megatron on the training side, SGLang on rollout)\n\nOver the past month, the team has been adding to this stack: more task environments, more environment types, and more training time. The result shows that \"environment scale for capability\" is a very efficient route.\n\n## Coding: The Longer the Task, the Bigger the Jump\n\nGLM-5.3's gains on coding benchmarks follow a clean pattern: **the longer the task, the larger the improvement**.\n\nKey numbers:\n\n- **Terminal-Bench 3.0**: 4.6 → 28.3 (over 6x)\n- **DeepSWE v1.1**: 46.2 → 66.9\n- **Agents' Last Exam (CLI variant)**: 23.8 → 28.5\n- **FrontierSWE**: 67.5 → 78.1\n- **GDPval-AA v2** (covering 44 occupations): 1508 → 1769\n\nOn Z.ai's own Code Bench, GLM-5.3 improves 50% over GLM-5.2. At Max effort, it hits 34.5% at roughly 75K output tokens per task, versus GLM-5.2's 23.4% at 96K tokens. At High effort, GLM-5.3 hits 31.4% at around 50K tokens — **beating Claude Opus 4.8's 29.5% at 120K tokens with fewer tokens**.\n\nClaude Fable 5 still leads at Max effort (39.5%), and GPT-5.6 Sol still edges out GLM-5.3 on Terminal-Bench 3.0 (34.6% vs 28.3%).\n\n## Cybersecurity: An \"Emergence\" That Surprised Z.ai Itself\n\nThis is the part of GLM-5.3 that genuinely surprised the industry.\n\nZ.ai openly admits in the blog: they only wanted to make the model better at finding vulnerabilities. They added vulnerability-discovery data and environments to the training mix. **The result went further than that**.\n\n> \"What surprised us was how quickly the capability continued to develop as training scaled.\" — Z.ai official blog\n\nConcrete benchmarks:\n\n- **CyberGym** (white-box vulnerability discovery): 77.2% → **84.5%**, ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%) — first place\n- **ExploitBench** (requires deeper root-cause reasoning and a working exploit): 24.4% → **54.4%**, more than doubling\n- **ExploitGym** (tasks completed under time-normalized budgets): within 2 hours, 29 → **105**; within 6 hours, 39 → **130**\n- Mythos 5 still leads on ExploitBench (78.0%) and on ExploitGym (181 \u002F 247 tasks)\n\nZ.ai's own framing: **the deeper into the exploitation chain a benchmark sits, the larger the improvement from GLM-5.2 — and also the wider the remaining gap to closed frontier models**. In other words, the model did not just find bugs — it began **forming coherent plans across complete exploitation chains**.\n\nThe most telling observation: **capability is growing fastest exactly where we are furthest behind**.\n\n## Real-World Validation: 269 Projects, 2,436 Vulnerabilities, One from 1981\n\nTo verify these capabilities were not overfitting to benchmarks, Z.ai has worked with several Chinese security teams since GLM-5.2 to run the model against real-world codebases.\n\nAfter expert review, screening, and deduplication:\n\n- 269 open-source projects\n- 2,436 vulnerability findings\n- 1,097 of them medium-to-high severity\n- Coverage: system kernels, operating systems, browser engines, open-source infrastructure, web applications, network protocols\n- **The oldest finding dates back to 1981**; on average, each vulnerability lived 26.6 years before discovery\n\nZ.ai built the [Security Disclosure Ledger](https:\u002F\u002Fcvd.z.ai\u002F) to track these findings through the disclosure process in public. At time of writing: 53 publicly disclosed, 2,383 still under embargo.\n\n## slime: The Plumbing Behind the 2.3× Training Throughput\n\nAll the above runs on Z.ai's open-source [slime](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime) framework. The design keeps training, rollout, and the data buffer on a single dataflow, so adding new RL environments becomes \"data generation\" rather than rewriting the training loop.\n\nTwo upgrades shipped alongside GLM-5.3:\n\n1. **Algorithmic**: top-p mask, top-k and full-vocabulary OPD, R3-style configurations improving training-rollout consistency. Average logprob difference is now controlled at the **1e-7 level** (>99.99% reduction vs. previous setups).\n2. **System**: local storage as an additional caching layer; multi-teacher OPD no longer requires long-running inference services per teacher; long-horizon coding RL training throughput improved by **over 2.3×** end-to-end.\n\nIn short: GLM-5.3's capability is not from a new base — it is from \"environments + algorithms + systems\" being scaled together.\n\n## How to Use It + One Gotcha\n\n- Live on the [Z.ai API](https:\u002F\u002Fdocs.z.ai\u002Fguides\u002Fllm\u002Fglm-5) and GLM Coding Plan\n- Works with Claude Code, OpenCode, and other major coding agent harnesses\n- Weights public in two weeks, after safety evaluation and hardening\n- **API change**: GLM-5.3 no longer supports ; it is unified to  + . Requests with the old config will fail\n- GLM Coding Plan switched to a points-based quota; 14:00–18:00 (UTC+8) on weekdays is peak hours, all other times get a 50% off-peak discount\n\n## Personal Commentary\n\nThe biggest question GLM-5.3 leaves the industry is not \"how strong is GLM-5.3?\" It is:\n\n1. **Is post-training scaling still showing diminishing returns?** On the same 743B base, Gym-class benchmarks moved this much in two months — that means our prior assumption that \"base model capability defines the ceiling\" may be wrong. The RL-environment scale and quality curve is still steep.\n2. **The security paradox of emergent cyber capability**: the better these models get at finding vulnerabilities, the more justified Anthropic's strict release controls on Mythos 5 become. Z.ai chose to ship online first, then open weights two weeks later, and built a public disclosure ledger as a first-class citizen — if this path works, it may be the first time an open-weight model has put \"disclosure workflow\" at the top of its release checklist.\n3. **The depth-curve of open vs. closed on cyber**: Z.ai's own observation — \"capability grows fastest where the gap is widest\" — means the open\u002Fclosed gap on cyber will not naturally converge; it will redistribute by how close tasks sit to real exploitation.\n\nIn the short term, this is a clear signal for teams building application security, CI remediation, or code auditing tools: long-horizon agents are already near closed-frontier capability on benchmarks, and the productization window may be the next 2-3 months.\n\nIn the longer view, the biggest beneficiary of this release may not be any single model vendor — it is everyone who already uses GLM-5.2 in production. The composability of post-training means we can now repeatedly push a \"good enough\" base to new positions through RL environments, without burning a fresh pre-training run each time.\n\n---\n\nReferences:\n- [GLM-5.3 Official Technical Blog](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3)\n- [MarkTechPost Coverage](https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F08\u002F14\u002Fz-ai-ships-glm-5-3-without-retraining-the-base-model-better-at-complex-coding-and-long-horizon-tasks\u002F)\n- [slime Open-Source Framework](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime)\n- [Z.ai Security Disclosure Ledger](https:\u002F\u002Fcvd.z.ai\u002F)","glm-5-3-post-training-emergent-cyber","2026-08-14T08:00:00Z","2026-08-14T09:06:22.091800Z","2026-08-14T09:06:22.091809Z",true,"agent",681,[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},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"de219584-58fc-45ad-91ea-0049a5cbcf10","OpenAI 开源 Codex Security CLI:把 AI 安全检测塞进每个 PR","openai-codex-security-cli-opensource","2026-07-29T10:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"1e7d0673-aecc-42b5-8560-92a2b4d4daf6","快手 KAT-Coder-V2.5 把 Agentic Coding 训练改写成基础设施工程","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"95b38793-01e2-48d7-a02b-24bf02b2fed5","xAI 把 Grok Build 全栈开源：刚被曝偷传完整代码库的 coding agent，如何用 Apache 2.0 救场","xai-grok-build-open-source","2026-07-16T16:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"6ceaf229-f1a2-4231-b531-797a99faa194","Ornith-1.0：模型自写 RL harness，SWE-Bench 比肩 Opus 4.7","ornith-1-0-397b-moe-swe-bench-opus-4-7","2026-06-26T18:01:01+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d63524e2-85bc-484c-b4e7-7fac32c3ac08","GLM-5.2 即将全量上线 Coding Plan：智谱把\"编程开源\"卷成新一轮标配","glm-5-2-coding-plan-zhipu-open-source","2026-06-13T07:30:00+00:00"]