20:"$Sreact.view_transition" 21:I[9754,["/_next/static/chunks/2t36f5p4pjzlb.js","/_next/static/chunks/1gyqeidjm5eoa.js","/_next/static/chunks/0_fjnxc4tt_l_.js","/_next/static/chunks/1ymt1shyhmu-4.js","/_next/static/chunks/3kv2zf0i7s2wc.js","/_next/static/chunks/1vs_3g_g75zmk.js","/_next/static/chunks/2g4dh802s8oz3.js","/_next/static/chunks/1bc3f0shukf1j.js","/_next/static/chunks/192kgs9v_t14j.js","/_next/static/chunks/0z6pw0v5c_lx5.js","/_next/static/chunks/08r60e6icmtkc.js","/_next/static/chunks/3trx6c0i7_x-3.js","/_next/static/chunks/40sro5bzp9bi_.js","/_next/static/chunks/1wgf3tk38y6cf.js"],"ProjectDetail"] 22:T46e4, GLM-5 · 从 Vibe Coding 到智能体工程

GLM-5 系列 · zai-org

从 Vibe Coding
到智能体工程

面向长时程智能体任务的开源大模型系列

把思考强度物化为一座可逐档棘轮上推的自主卡齿轨——档位越低越接近一句即答，档位越高越能自行走完长链条任务。GLM-5、GLM-5.1、GLM-5.2 每一代，都是往更高卡齿推进的一格。

查看仓库阅读技术报告

Apache-2.0 6,983 Stars 903 Forks 42 Open Issues zai-org/glm-5

思考强度：可逐档上推的卡齿槽

GLM-5.2 提供多个思考强度档位，用来在性能与延迟之间取得平衡。档位越低越接近爽利的即答，档位越高越接近能自行走完长链条任务的工程能力。

档位 · 低稳即答

低延迟，一句话进入状态，适合浅触与快速试探。

档位 · 中任务分解

多步推理与意图保持，在被上下文耗尽打断前走得更远。

档位 · 高长链条

可独立调度并推完长链条任务，接近智能体工程。

档位 · 顶自主可 1M

稳定的百万级上下文，为长航任务提供纵深。

三代站台依次拾级

每一代 GLM 都是卡齿轨上一格站台，用冲压数字与测标旗挂上各自的登顶基准。

GLM-5

第一档 · 登顶

#1 开源

Vending Bench 2开源模型第一 · 终局余额 $4,432

GLM-5.1

第二档 · 领先

SOTA

SWE-Bench Pro达先进水平

NL2Repo · Terminal-Bench 2.0大幅领先 GLM-5

GLM-5.2

第三档 · 深度

上下文稳定百万 token

思考强度多档可调平衡延迟

装配底盘 · 本地部署

GLM-5 系列支持以多套框架上架：把模型装到自己的底盘上，随时开工。

部署框架RUNTIME PLATE

SGLang

vLLM

Transformers

KTransformers

Unsloth

调校压机 · 微调

用 Slime 与 ms-swift 为模型施加 SFT、PPO、GRPO 等调校，把它压出你要的形态。

微调工具TUNER PRESS

工具Slime

工具ms-swift

方法SFT

方法PPO / GRPO

如实铭牌

规模与成绩都写清楚。以下是与部署、基准相关的真实边界。

资源门槛

模型体积大（744B 参数），部署资源要求较高。

官方自测

基准成绩来自官方自测，实际效果需自行验证。

框架依赖

部分能力（如思考强度控制）需特定框架支持。

23:T45ba, GLM-5 · From Vibe Coding to Agentic Engineering

GLM-5 Series · zai-org

From Vibe Coding
to Agentic Engineering

An open-source LLM series for long-horizon agentic tasks

Treat thinking effort as a ratchet you push up one notch at a time. The lower rungs answer on the fly; the higher rungs let the model carry a long chain of tasks to completion. Each generation of GLM-5 pushes a notch further up.

View the repository Read the report

Apache-2.0 6,983 Stars 903 Forks 42 Open Issues zai-org/glm-5

Thinking effort, rung by rung

GLM-5.2 exposes several thinking-effort levels so you can trade performance against latency. Lower detents answer briskly; higher detents push a long chain toward real engineering.

Detent · LowInstant

Low latency and one-shot answers, ideal for quick probes and light touches.

Detent · MidDecompose

Multi-step reasoning with intent held across turns before context runs out.

Detent · HighLong chain

Schedules and pushes long-horizon tasks to completion, closer to agentic engineering.

Detent · Top1M reach

A steady 1M-token context supplies the depth long tasks need.

Three generations, three rungs

Each GLM generation is a station on the rack, its summit benchmarks hung as stamped numbers and flags.

GLM-5

Rung one · summit

#1 OSS

Vending Bench 2#1 open-source · ends at $4,432

GLM-5.1

Rung two · lead

SOTA

SWE-Bench Prostate-of-the-art

NL2Repo · Terminal-Bench 2.0well ahead of GLM-5

GLM-5.2

Rung three · depth

Contextsteady million-token window

Thinking effortmultiple levels to tune latency

Assembly plate · Deploy locally

The GLM-5 series mounts on a range of runtimes, so the model slots onto your own hardware and gets to work.

Deployment frameworksRUNTIME PLATE

SGLang

vLLM

Transformers

KTransformers

Unsloth

Tuner press · Fine-tuning

Apply SFT, PPO, and GRPO through Slime and ms-swift to press the model into the shape you need.

Tuning toolingTUNER PRESS

ToolSlime

Toolms-swift

MethodSFT

MethodPPO / GRPO

An honest nameplate

Scale and scores written plainly — the real edges around deployment and benchmarks.

Resource cost

Large model size (744B params) asks for significant deployment resources.

Self-reported

Benchmark figures are officially self-reported; verify them in your own use.

Framework-bound

Some features (such as thinking-effort control) require a specific runtime.

a:[["$","$20",null,{"enter":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-enter","default":"none"},"exit":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-exit","default":"none"},"update":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-update","default":"none"},"default":"none","children":["$","$L21","cmt2zh451023u6271n26u27qb:zh-CN:beautified",{"project":{"id":"cmt2zh451023u6271n26u27qb","slug":"zai-org-glm-5","name":"glm-5","description":"面向长时程智能体任务的开源大模型系列","authorName":"zai-org","authorVerification":{"verified":false,"verifiedAt":null},"category":"development_tools","subcategory":"data_ai_tools","experienceUrl":"https://z.ai/blog/glm-5.2","qualityScore":88,"featured":false,"officialSite":null,"views":{"original":{"available":true},"official":{"available":false,"reason":"disabled_by_policy"},"beautified":{"available":true,"reason":null},"defaultView":"beautified"},"beautifiedView":{"revisionId":"cmt2zh45g02466271mlvethgz","kind":"artifact","artifact":{"schemaVersion":"localized-static-html-v4","kind":"static_html","locales":["zh-CN","en"],"defaultLocale":"zh-CN","documents":{"zh-CN":"$22","en":"$23"},"designDirections":{"schemaVersion":"dodot-design-directions-v1","directions":[{"id":"context-horizon-window","name":"Context Horizon Window","projectMetaphor":"把「单一任务能走多远」视为一扇随模型理解范围而张开的观测窗：弦开得越宽，一个长时程智能体任务在被上下文耗尽打断前能到达的纵深越远。核心张力是「1 turn 的浅触」与「100 万 token 的长航」之间的跨度。","signatureVisual":"无照片、暖底绘图纸上的扩展弧量规：从左侧「1 turn」小手笔出发，一串同心的海波状弧逐渐展开到右侧大半径处落在「1M tokens」刻度；三个模型版本 GLM-5 / 5.1 / 5.2 作为弧线上不同高度的刻渡标，每个刻渡标旁以印章式数据标签压上它延伸到的能力坐标。手工制图线、量角刻度，不做任何发光。","heroGeometry":"页面顶部一张大跨度的弧形圆规扫掠：圆心落在视口下方深处，同心的刻度环带向上辐射，形成向左右扩展的「观测篮窗」——将横幅结构转化为一枚可读的能力半径罗盘。","sectionSequence":["horizon-hero：能力半径罗盘 + 一句话价值 + 真实 CTA","arc-benchmark-ascent：SWE-Bench Pro / NL2Repo / Terminal-Bench / Vending Bench 2 沿弧线高处依次登攀的刻渡标","runtime-reach：SGLang / vLLM / Transformers / KTransformers / Unsloth 五个运行时托盘的接入半径","fine-tune-dial：Slime / ms-swift（SFT / PPO / GRPO）的调教回转刻度","honest-capacity：744B 参数与官方自测的如实备注列"],"surfaceModel":"暖米色工程绘图纸：细头发丝制图线、量角刻度环、印章式数据标签；克制的墨水密度，单一强调色用于弧线本体，所有坐标与刻度都是真实证据数字，无渐变光晕。","palette":"暖图纸底 #F5F0E6，墨水主色 #1F241A，强调弧线 #B5542E，量角刻度 #6B7263，印章红点 #C0392B 用于自测警告。","typography":"制图排字风：标题用几何无衬线大号字形，数据标签用等宽（仅数字坐标），正文用中等字重的无衬线。中英各自语感，字号不追求完全对等。","desktopComposition":"1280 宽：左上定锚一句话价值，右侧为展开的弧线量规作为唯一图形焦点；五个章节在圆弧不同半径处错落排布，刻度与印章沿弧分布，而非并排三卡。","mobileTransformation":"390 宽：弧度坍缩为右侧一条垂直的「上下文含量尺」——弯弓变成一根随半径读数延伸的竖直标尺，弧线上的刻渡标重排为竖向堆叠的测量条目，量程自上而下读取。","mediaStrategy":"assetManifest 为空，无本地合格素材；删除一切媒体区，以证据驱动的弧线量规与印章刻度构成标志视觉，不造假产品截图，不占位空框。","forbiddenPatterns":["拒绝三张指标卡并排","拒绝浏览器 / 终端 / 编辑器壳与伪造界面控件","拒绝泛用渐变光晕与粒子背景，不把光泽当视觉想法","拒绝营销横幅式大标题占位"],"compositionFingerprint":"arc-range-gauge:1-turn-left-scope-sweeping-to-1M-token-radius-with-three-model-altitude-ticks-and-stamp-data-labels-on-matte-drafting-paper"},{"id":"autonomy-effort-rack","name":"Autonomy Effort Rack","projectMetaphor":"把「思考强度」物化为一座可逐档棘轮上推的自主卡齿轨：档位越低越接近一句即答的 vibe coding，档位越高越接近能自行走完长链条任务的智能体工程；每一代 GLM（5 → 5.1 → 5.2）都是往更高卡齿推进的一格。","signatureVisual":"竖直的棘齿卡轨自下而上堆叠：每个卡齿槽映射一档思考强度，顶部卡齿是一块锚定的「Agentic Engineering」楔石；三个模型版本作为相邻卡齿上的站台，站台面用冲压数字与测标旗挂上各自登顶的基准（SWE-Bench Pro / NL2Repo / Terminal-Bench / Vending Bench 2）。无照片，车间造物质感，靠实体块的堆叠与冲压字表达，不做任何发光。","heroGeometry":"视口左侧一根贯穿全高的竖直卡齿轨，上方数份卡齿逐级升高，顶端以「Agentic Engineering」为封顶楔石；右侧紧贴冲压式数值标签。单杆承载的楼梯式自主度刻度，不同于横幅也不做网格。","sectionSequence":["rack-hero：vibe coding → agentic engineering 的自主度棘轮","effort-detents：多档思考强度与延迟的卡齿档位映射","generation-rungs：GLM-5 / 5.1 / 5.2 三代站台依次拾级","deploy-plate：SGLang / vLLM / Transformers / KTransformers / Unsloth 的装配底盘","tuner-press：Slime / ms-swift（SFT / PPO / GRPO）的调校压机","caveat-plate：744B 参数与自测成绩的如实铭牌"],"surfaceModel":"车间金属板：冲压数字、实体块堆叠、铆点式小圆点、细格栅测度；金属冷灰为骨、单一强调色用于最上档自主度楔石，实体与留白交替，无阴影镂空滥用。","palette":"冷灰金属底 #E7E4DE，铁灰 #23272E，齿槽深灰 #4A4F57，自主度强调 #C1502E，冲压数字 #1B1F24，铆点 #8A8578。","typography":"车间铭牌风：标题用高对比几何字形，档位号与数字用宽等宽冲压字样，正文中等无衬线；中英自然成句。","desktopComposition":"1280 宽：竖直卡齿轨独占左侧一列通栏，第三代站台、部署底盘、调校压机与铭牌作为右侧纵向堆叠的工作台序列，主次分明、不用卡片墙。","mobileTransformation":"390 宽：竖直卡齿轨旋转为一条横向的进度走棘——由左至右逐档推进，档位卡齿变成可滑动的堆叠条目，楔石落在右侧终点，自主度读数横向读取。","mediaStrategy":"assetManifest 为空，无本地合格素材；删除媒体区，以卡齿轨、冲压数字与测标旗构成证据驱动的标志视觉，不造假截图、不占位框。","forbiddenPatterns":["拒绝三卡指标并排","拒绝浏览器 / 编辑器 / 终端壳与伪造开关滑块","拒绝泛用发光网格与粒子，光泽不做视觉替身","拒绝把桌面三列在手机上原样竖堆"],"compositionFingerprint":"autonomy-effort-rack:single-vertical-notch-ratchet-escalating-from-vibe-coding-to-a-keystoned-agentic-engineering-with-thinking-effort-detents-and-generation-stations-as-stamped-rungs"},{"id":"split-order-column","name":"Split Ordered Column","projectMetaphor":"把「从 Vibe Coding 到 Agentic Engineering」表现为同一条焊缝两侧的剖面对照：焊缝左侧是未经编排的随机碎屑（一句即答、不连贯步进、无状态），焊缝右侧是编排成序的工程层；模型的 1M 上下文与思考强度是把左侧碎屑逐块拼接并卡入右侧空槽的编排机制。","signatureVisual":"整页被一道竖缝一分为二：左列是一团松散散落的无序碎块（以抽象几何拼块与断裂短线示意无序，非任何真实 UI），右列是层层严整堆叠的工程层；一根接续线从每个碎块所在处引出、穿过缝，落进右侧对应层槽，表达「编排把他们整理归位」。墨线单色为主、唯一强调色用于接续线与缝本身，手工剖面对照图，无发光。","heroGeometry":"贯穿全页的镜像不对称双列柱与中央一道竖直接缝线把横幅劈作左右两半，左「乱」右「序」，接缝处是唯一强调色的连续性来源。","sectionSequence":["split-hero：无序碎屑与有序工程层的剖面对照","disorder-ledger：无序来源的清单（示意图式，不冒充截图）","ordering-mechanism：长上下文与思考强度如何把碎屑拼成层的机理","benchmark-reconcile：基准成绩与「官方自测需自行验证」的对账","runtime-appendices：部署框架与微调工具两类附件","limits-column：744B 资源门槛与待验证事项的诚实栏"],"surfaceModel":"平版印刷正交柱体：细发丝对准线、断面填充纹理、单色墨为主、一处强调色专用于接缝与接续线；全部由真实证据驱动的对齐关系构成，无渐变与投影堆砌。","palette":"白底 #FBFBF8，墨水 #232323，断面灰 #D9D6CE，接缝强调 #C1502E，注记铁灰 #7A766C。","typography":"工程蓝图排字风：标题用超细 + 对比性粗体对置，序号与对准值用等宽，正文常规无衬线；中英各自密度，不追求逐字对译。","desktopComposition":"1280 宽：左右双柱固定在视口内对称剖开，接缝竖直；六章节若分布在柱内上下错位，章节本身以接缝线为主导贯穿，避免并列卡墙。","mobileTransformation":"390 宽：竖直双柱与接缝旋转 90°——变为上方一段「乱」带与下方一段「序」带，中央接缝成为水平分界；接续线重排为纵向下沉的衔接链，自乱入序逐条对接。","mediaStrategy":"assetManifest 为空，无本地合格素材；删除媒体区，以双柱剖面对照、接缝与接续线构成证据驱动的标志视觉，不造假截图、不占位框。","forbiddenPatterns":["拒绝三张指标卡并排","拒绝伪造 tabs / 开关 / 输入 / 图表控件","拒绝泛用粒子与光晕背景代替项目视觉","拒绝桌面双列在手机上仅简单竖排而不重组"],"compositionFingerprint":"split-order-column:full-height-asymmetric-double-column-bisected-by-a-single-seam-with-chaos-fragments-on-one-side-engineered-layers-on-the-other-and-dashed-relay-lines-routing-each-fragment-into-its-ordered-slot"}],"selectedDirectionId":"autonomy-effort-rack","selectionRationale":"目录中三个方向隐喻与几何均真正不同（Arc 展开量规 / 竖直棘轮卡轨 / 双柱剖面对照）。选「autonomy-effort-rack」是因为它最能表达项目机制：项目的标题命题就是「从 Vibe Coding 到 Agentic Engineering」，其关键实现机制恰是「思考强度档位」与「逐代推进的长时程能力」——把思考强度物化为可逐档上推的棘轮、让三代模型作为拾级站台，是对该机制最贴切、最直接的纯证据构图。同时 SC 基准（SWE-Bench Pro / NL2Repo / Terminal-Bench / Vending Bench 2）与部署/微调事实能以测标旗与铭牌自然挂载到齿轨上，几乎不存在造图成分；其冷灰车间金属表面亦区别于 anti-template 中多见的暖图纸、流线、波动与反应循环构图。综合项目契合、证据契合、标志视觉辨识度与相似风险最低，选定该方向。"},"designBrief":{"schemaVersion":"dodot-design-brief-v1","selectedDirectionId":"autonomy-effort-rack","projectMetaphor":"把「思考强度」物化为一座可逐档棘轮上推的自主卡齿轨：档位越低越接近一句即答的 vibe coding，档位越高越接近能自行走完长链条任务的智能体工程；每一代 GLM（5 → 5.1 → 5.2）都是往更高卡齿推进的一格。","signatureVisual":"竖直的棘齿卡轨自下而上堆叠：每个卡齿槽映射一档思考强度，顶部卡齿是一块锚定的「Agentic Engineering」楔石；三个模型版本作为相邻卡齿上的站台，站台面用冲压数字与测标旗挂上各自登顶的基准（SWE-Bench Pro / NL2Repo / Terminal-Bench / Vending Bench 2）。无照片，车间造物质感，靠实体块的堆叠与冲压字表达，不做任何发光。","compositionFingerprint":"autonomy-effort-rack:single-vertical-notch-ratchet-escalating-from-vibe-coding-to-a-keystoned-agentic-engineering-with-thinking-effort-detents-and-generation-stations-as-stamped-rungs","heroGeometry":"视口左侧一根贯穿全高的竖直卡齿轨，上方数份卡齿逐级升高，顶端以「Agentic Engineering」为封顶楔石；右侧紧贴冲压式数值标签。单杆承载的楼梯式自主度刻度，不同于横幅也不做网格。","sectionSequence":["rack-hero：vibe coding → agentic engineering 的自主度棘轮","effort-detents：多档思考强度与延迟的卡齿档位映射","generation-rungs：GLM-5 / 5.1 / 5.2 三代站台依次拾级","deploy-plate：SGLang / vLLM / Transformers / KTransformers / Unsloth 的装配底盘","tuner-press：Slime / ms-swift（SFT / PPO / GRPO）的调校压机","caveat-plate：744B 参数与自测成绩的如实铭牌"],"surfaceModel":"车间金属板：冲压数字、实体块堆叠、铆点式小圆点、细格栅测度；金属冷灰为骨、单一强调色用于最上档自主度楔石，实体与留白交替，无阴影镂空滥用。","palette":"冷灰金属底 #E7E4DE，铁灰 #23272E，齿槽深灰 #4A4F57，自主度强调 #C1502E，冲压数字 #1B1F24，铆点 #8A8578。","typography":"车间铭牌风：标题用高对比几何字形，档位号与数字用宽等宽冲压字样，正文中等无衬线；中英自然成句。","desktopComposition":"1280 宽：竖直卡齿轨独占左侧一列通栏，第三代站台、部署底盘、调校压机与铭牌作为右侧纵向堆叠的工作台序列，主次分明、不用卡片墙。","mobileTransformation":"390 宽：竖直卡齿轨旋转为一条横向的进度走棘——由左至右逐档推进，档位卡齿变成可滑动的堆叠条目，楔石落在右侧终点，自主度读数横向读取。","mediaStrategy":"assetManifest 为空，无本地合格素材；删除媒体区，以卡齿轨、冲压数字与测标旗构成证据驱动的标志视觉，不造假截图、不占位框。","evidencePlan":["仓库事实：Apache-2.0、6983 stars、903 forks、442 open issues、zai-org/glm-5，作为铭牌/底盘部分的冲压元数据。","基准证据：GLM-5.1 在 SWE-Bench Pro 达先进水平，且于 NL2Repo、Terminal-Bench 2.0 领先 GLM-5；GLM-5 于 Vending Bench 2 开源第一、模拟一年售货机业务最终余额 $4,432——作为各代站台的测标旗数据，并在 caveat 栏注明官方自测需自行验证。","能力事实：GLM-5.2 提供稳定 1M token 上下文与多档思考强度；本地部署框架 SGLang / vLLM / Transformers / KTransformers / Unsloth，微调 Slime / ms-swift（SFT / PPO / GRPO），作为档位、部署底盘与调校压机的内容来源。","诚实栏：744B 参数部署资源要求高、基准为官方自测、部分功能需特定框架支持——作为 caveat-plate 的如实铭牌。"],"localeContract":{"zh-CN":"从 Vibe Coding 到智能体工程——面向长时程智能体任务的开源大模型系列","en":"From Vibe Coding to Agentic Engineering — an open-source LLM series for long-horizon agentic tasks"},"forbiddenPatterns":["拒绝三卡指标并排","拒绝浏览器 / 编辑器 / 终端壳与伪造开关、滑块、tabs、搜索、输入等控件","拒绝泛用发光网格、粒子或渐变光晕作为视觉替身","拒绝把桌面三列在手机上原样竖堆而不重组","拒绝 fake 客户 LOGO、评测图表或产品截图占位"]},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"approved","summary":"zh-CN.html 与 en.html 忠实实现了项目专属设计简报「autonomy-effort-rack」：竖直棘齿卡轨（vibe coding → agentic engineering）作为贯穿视口左侧的标志视觉，顶部「智能体工程 / AGENTIC ENGINEERING」强调色楔石封顶，档位卡齿逐级升高并带冲压编号与测标旗；六个章节严格对应 sectionSequence（rack-hero/effort-detents/generation-rungs/deploy-plate/tuner-press/caveat-plate）。所有基准、能力、部署/微调与诚实栏数据均以 material.json 事实为依据（Apache-2.0、6,983 stars、903 forks、42 open issues、SWE-Bench Pro / NL2Repo / Terminal-Bench 2.0 / Vending Bench 2、$4,432、1M token、744B 参数、官方自测等）。手机 390 断点处卡齿轨翻转重构成横向走棘、档位变为堆叠条目、楔石落在右侧终点，形成真正重组而非竖堆。双语章节语义等价且各语自然成句。无通用 Hero、无三卡指标并排、无设备壳、无装饰渐变网格（铆点质感为简报 surfaceModel 的组成部分），无假控件（CTA 为真实链接，全站 fakeControlCount=0）。七项检查全部通过，无 blocking issue。","checks":{"projectFit":{"passed":true,"evidence":"六章节顺序与设计简报 sectionSequence 完全一致：rack-hero → effort-detents → generation-rungs → deploy-plate → tuner-press → caveat-plate（zh/en 均以 data-dodot-section 标注）。色板精确匹配简报 palette（#E7E4DE 底、#23272E 铁灰、#4A4F57 槽灰、#C1502E 强调、#1B1F24 冲压、#8A8578 铆点）。桌面 1280 下竖直卡齿轨独占左侧 360px 通栏，右侧为纵向堆叠的工作台序列，主次分明，符合 desktopComposition，未用卡片墙。"},"evidenceBound":{"passed":true,"evidence":"仓库元数据 Apache-2.0 / 6,983 stars / 903 forks / 42 open issues / zai-org/glm-5 来自 material.json repository 字段（real，且 openIssues=42 以原始材料为准）。三代站台分别挂载 Vending Bench 2 #1 开源+终局余额 $4,432（对应 claim4）、SWE-Bench Pro 先进+NL2Repo/Terminal-Bench 2.0 领先（claim3）、1M token+多档思考强度（claim1/2）。部署框架 SGLang/vLLM/Transformers/KTransformers/Unsloth（claim5）、微调 Slime/ms-swift 含 SFT/PPO/GRPO（claim6）属实。铭牌如实含 744B 参数、官方自测需自行验证、功能需特定框架支持——与 caveats 一致。"},"signatureVisual":{"passed":true,"evidence":"标志视觉真实存在于 HTML：`.rack` 内自下而上堆叠四个递增卡齿 `.notch-1..4`（高度 64→78→92→106px、margin-left 逐档收窄形成楼梯式自主度刻度），顶部 `.keystone` 以独立强调色承载「智能体工程 / AGENTIC ENGINEERING」封顶楔石，带 `rivet-grid` 铆点烘焙质感，冲压编号 + 右侧数值标签；铭牌/顶梯度挂测标旗用冲压数字与 flag 表达 SC 基准。纯几何、无照片、不做发光，忠实于 signatureVisual 描述与 surfaceModel 车间造物观感。"},"mobileTransformation":{"passed":true,"evidence":"媒体查询 max-width:760px 将 `.rack-rail` 由纵向 flex 翻转 `flex-direction:row`，卡齿变为横向堆叠条目（flex-basis 34%→46%→60%→74% 递增），`.keystone` 收窄为 88% 宽并落在右侧终点，导轨 `.rails` 旋转为横向走棘，自主度读数改为横向读取；同时 worktable/gen-col/chip-grid/split-grid/caveat-grid 各自折叠为单列或两列。这是真正的旋转重组而非桌面三列原样竖堆，符合 mobileTransformation。"},"localeParity":{"passed":true,"evidence":"zh-CN 与 en 六章节语义一一对应且各自自然成句：标题『从 Vibe Coding 到智能体工程 / From Vibe Coding to Agentic Engineering』、hero-lede 卡齿比喻、four detents、三代站台 flag 数据、部署芯片、调校压机与铭牌均等价；译文非逐字对译（如 Detent·Low=档位·低、Instant=稳即答），各自语感成立，数字与事实一致。"},"antiTemplate":{"passed":true,"evidence":"构图以竖直卡齿轨标题视觉 + 右侧冲压档案/底盘/压机/铭牌构成，非通用 Hero + 三卡 + 卡片墙；支架不落在三卡指标并排（基准以三代站台内嵌 flag 呈现，仓库元数据置于 meta-strip 而非指标卡）。无浏览器/编辑器/终端设备壳，无泛用发光网格或粒子当视觉替身。铆点圆点属简报 surfaceModel 明示的材质纹理（非装饰性渐变网格），未滥用。"},"noFakeUi":{"passed":true,"evidence":"全站无伪造开关/滑块/tabs/搜索/输入/图表控件；两个 CTA 均为真实 ``（github.com/zai-org/glm-5 与 z.ai/blog/glm-5.2），assetManifest 为空、无本地媒体故按 mediaStrategy 删除媒体区也未放占位框。brQA fakeControlCount=0、emptyLinkCount=0、brokenImageCount=0，无假 LOGO 或产品截图占位。"}},"issues":[]},"repairEvidence":[],"provenance":{"executor":"opencode","model":"deepseek/deepseek-v4-flash","opencodeVersion":"1.1.58","promptVersion":"dodot-showcase-deepseek-v4","inputHash":"814296097b68982596e20b4374291c9cc3dc3cdefe9769f6ee0f10c52a6fd7fb","artifactHash":"5c737ecf294f15370be08185344bd7a4f5cf27431f737b0ed0b76b00ba3cd643","designSkill":"dodot-showcase-design"},"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":5,"sectionCount":6,"textLength":1200,"cjkCount":578,"latinLetterCount":283,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":3528,"h1Top":342.390625,"bottomInset":152.390625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"421c6332491ca328352d3015b07bc6feeff92740abc6b5b3e621e8f5674dd28d","middle":"5dcfc6573144d34d2d61897520d6b50de635dfab194c51542b60a9d9b806e00b","bottom":"bf837e61368152d384aabe2a447aecab071694d25e9fe811326c9c627cec4016"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":5,"sectionCount":6,"textLength":1200,"cjkCount":578,"latinLetterCount":283,"scrollWidth":390,"clientWidth":390,"scrollHeight":4444,"h1Top":381.1875,"bottomInset":120.25,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"f1d1cca2b55edf8c36bac71137934bdc35d47de7b717876dac22d4e1c656a875","middle":"9832fe5b41ee34ead19f39675f20c9a02aa826f45fe0d1360e2a7e28123971ca","bottom":"78909bc4abd1722afdcc55a0798fbefc0021b901b27ef1a308089ad6570984b4"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":5,"sectionCount":6,"textLength":2519,"cjkCount":0,"latinLetterCount":1938,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":3646,"h1Top":300.234375,"bottomInset":152,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"771e1757f5dae9b6f113aa3b6184e2504debd8d9dc4c60d9591a5e86932cf731","middle":"65a7267972388b70b5d552c76c35601392f8f52f38b54f76f8a026d70a2e8417","bottom":"8ed17033eec0cc489e01ff17966b9dba204c3eb64baec21d365c8432c4c3fc85"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":5,"sectionCount":6,"textLength":2519,"cjkCount":0,"latinLetterCount":1938,"scrollWidth":390,"clientWidth":390,"scrollHeight":4759,"h1Top":381.1875,"bottomInset":120.390625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"90cb03520fe3bac4dde865382339d4c7930dc24cc94c3e71ede76d26c70ca6d0","middle":"13a15696f70002b09a1c0f1aa45fcdca58be8e61fa9713ee04132a294ceb04b2","bottom":"414c4da05a8e4eb173c3f4c4e9f2d41a7e6d494f0ebbffec0f1c7830a0b31152"},"consoleErrors":[],"networkRequests":[]}]}},"generatorVersion":"opencode-1.1.58:deepseek/deepseek-v4-flash:dodot-showcase-deepseek-v4","publishedAt":"2026-08-20T16:20:09.269Z"},"rawExtraction":{"projectRevisionId":"cmt2zh45602406271rfxnwsnf","projectVersion":1,"sources":[{"canonicalKey":"github:zai-org/glm-5","role":"primary","revisionVersion":1,"url":"https://github.com/zai-org/GLM-5"}],"facts":[{"key":"repository_owner","value":"zai-org","evidence":[{"path":"repository.owner","excerpt":"zai-org"}]},{"key":"repository_name","value":"glm-5","evidence":[{"path":"repository.name","excerpt":"glm-5"}]},{"key":"description","value":"GLM-5: From Vibe Coding to Agentic Engineering","evidence":[{"path":"repository.description","excerpt":"GLM-5: From Vibe Coding to Agentic Engineering"}]},{"key":"license","value":"Apache-2.0","evidence":[{"path":"repository.license","excerpt":"Apache-2.0"}]},{"key":"topics","value":"agentic-ai, coding, llm, long-horizon","evidence":[{"path":"repository.topics","excerpt":"agentic-ai, coding, llm, long-horizon"}]},{"key":"source_count","value":"1","evidence":[{"path":"projectRevision.sources","excerpt":"github:zai-org/glm-5"}]}],"taxonomy":{"category":"development_tools","subcategory":"data_ai_tools","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"GLM-5: From Vibe Coding to Agentic Engineering"}]},"lifecycle":{"stage":"prototype","confidence":0.65,"evidence":[{"path":"readmeExcerpt","excerpt":"# GLM-5.2 & GLM-5.1 & GLM-5

👋 Join our Wechat or Discord community.
📖 Check out the GLM-5.2 blog and GLM-5 Technical report.
📍 Use GLM-5.2 API services on

从 Vibe Coding到智能体工程