What Is GLM-5?GLM-5 是什么？
GLM-5 is Z.ai's (Zhipu AI) flagship open-weight model, released February 11, 2026 — the model Z.ai frames as a shift from "vibe coding" to agentic engineering: end-to-end project construction, refactors, deep debugging and multi-step tool workflows with tight goal consistency, not just snippet generation. The scale: 744 billion total parameters with about 40B active per token, trained on 28.5 trillion tokens (up from 23T on the previous generation), with a 200K-token context window. It was reportedly trained entirely on Huawei Ascend chips — a first for a frontier-class model — and released under the MIT license in BF16 and FP8 builds.GLM-5 是 Z.ai（智谱）2026 年 2 月 11 日发布的开权重旗舰，Z.ai 把它定位为从“氛围编程”到 Agent 工程的转向：端到端建项目、重构、深度排障、多步工具链路且目标一致性强，而不只是生成代码片段。体量：744B 总参数、每 token 激活约 40B，预训练 28.5T token（上一代 23T），200K 上下文。据报它完全在华为昇腾芯片上训练——前沿级模型里的首例——采用 MIT 许可，提供 BF16 与 FP8 版本。
Version discipline matters on this page: GLM-5 is not GLM-4.7 (the December 2025 predecessor), and its scores are not the later GLM-5.2 (June 2026) or GLM-5.3 (August 2026) numbers. GLM-5's own vendor-reported results: 77.8% on SWE-bench Verified — the open-weight record at launch — and 56.2 on Terminal-Bench 2.0, with an open-model #1 ranking in text arena comparisons at launch. All benchmark claims are vendor-reported.本页的版本纪律很重要：GLM-5 不是 GLM-4.7（2025 年 12 月的前代），它的分数也不是后来 GLM-5.2（2026 年 6 月）或 GLM-5.3（2026 年 8 月）的数字。GLM-5 自己的厂商自报成绩：SWE-bench Verified 77.8%——发布时开权重纪录——Terminal-Bench 2.0 56.2，发布时文本竞技场比较中开模型第一。基准数据一律为厂商自报。
GLM-5 matters less for any single benchmark than for what it proved: a frontier-class open model trained without Nvidia hardware, shipped under MIT, and immediately usable for real engineering work rather than demos. The agentic-engineering framing is the right lens — its value shows up in long task chains that stay on goal, including a stronger "abstain" behavior that third-party commentary highlighted, rather than in one-shot code completion. The honest limits: 200K context trails the 1M class, self-hosting a 744B model is a multi-GPU project even in FP8, and the February scores are already superseded by GLM-5.2/5.3. Verdict: adopt GLM-5 if you want the MIT-licensed February baseline with a clean provenance story; start a new project by comparing it against its own successors, and use the API (prices per official site) before committing to the GPU bill.
GLM-5 的意义不在于某一项基准，而在于它证明的事：一个不用英伟达硬件训出来的前沿级开模型，以 MIT 许可发布，立刻能干真实工程活而不只是演示。Agent 工程这个定位是正确的观察角度——它的价值体现在目标不跑偏的长任务链上，包括第三方评论提到的更强的“拒答”行为，而不是一次性代码补全。真实局限：200K 上下文落后于 100 万级别；744B 模型即使 FP8 也是多卡工程；2 月的分数已被 GLM-5.2/5.3 超过。结论：想要 MIT 许可的 2 月基线、且看重出身故事，就选 GLM-5；新项目动手前先拿它和自己的后继版本比一遍，GPU 账单拍板前先走 API（价格以官网为准）。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use GLM-5?谁适合使用 GLM-5？
✓ Good Fit For适合以下场景
- Long-horizon agentic engineering: full projects, refactors, deep debugging长程 Agent 工程：整项目、重构、深度排障
- Teams that need MIT weights with a clean commercial story需要 MIT 权重、商用出身干净的团队
- Workloads that fit in a 200K context window能装进 200K 上下文的工作负载
✕ Not Ideal For不适合以下场景
- Agent loops that need 1M-token context — look at 1M-class open models需要 100 万上下文的 Agent 链路——请看 100 万级别的开模型
- Anyone who wants the latest GLM scores — GLM-5.2/5.3 already post higher numbers想要最新 GLM 分数的人——GLM-5.2/5.3 的数字已经更高
- Small teams expecting single-GPU self-hosting of a 744B model指望单卡自托管 744B 模型的小团队
Getting Started with GLM-5GLM-5 快速开始
1. Confirm the version first: GLM-5 (Feb 2026) — not GLM-4.7, not GLM-5.2/5.3 — before downloading weights先确认版本：GLM-5（2026-02）——不是 GLM-4.7，也不是 GLM-5.2/5.3——再下载权重2. Try the API / chat first; check Z.ai's official pricing page for the current rate — prices per official site先走 API / 对话入口试；现价查 Z.ai 官网价目——价格以官网为准3. Self-hosting: FP8 build via vLLM or SGLang with multi-GPU tensor parallelism自托管：FP8 版走 vLLM 或 SGLang，多卡张量并行Key Features核心功能
- 744B MoE, 40B Active744B MoE、40B 激活 — 744 billion total parameters with about 40B active per token — more than double the GLM-4.7 generation's scale.744B 总参数、每 token 激活约 40B——体量是 GLM-4.7 世代的两倍多。
- 200K-Token Context200K 上下文 — 200K-token context window, served efficiently with DeepSeek-style sparse attention (DSA).200K 上下文，DeepSeek 式稀疏注意力（DSA）让长上下文服务成本可控。
- Agentic Engineering DesignAgent 工程设计 — Built for complex systems engineering and long-horizon agent tasks, with slime asynchronous RL for faster post-training iteration.为复杂系统工程与长程 Agent 任务而生，slime 异步强化学习加快后训练迭代。
- MIT License, BF16 / FP8 WeightsMIT 许可、BF16 / FP8 权重 — MIT-licensed weights in BF16 and FP8 builds — FP8 is the practical serving choice.MIT 许可权重，提供 BF16 与 FP8 版本——FP8 是实际服务时的务实选择。
- 77.8% SWE-bench VerifiedSWE-bench Verified 77.8% — Vendor-reported 77.8% on SWE-bench Verified (open-weight record at launch) and 56.2 on Terminal-Bench 2.0.厂商自报 SWE-bench Verified 77.8%（发布时开权重纪录）、Terminal-Bench 2.0 56.2。
Pros & Cons优缺点
✓ Pros优点
- Open-weight record at launch: 77.8% SWE-bench Verified (vendor-reported)发布时开权重纪录：SWE-bench Verified 77.8%（厂商自报）
- MIT license with BF16/FP8 weights — clean commercial useMIT 许可 + BF16/FP8 权重——商用干净
- Agentic reliability: goal-consistent long task chains, stronger abstain behaviorAgent 可靠性：长任务链目标一致、拒答行为更强
- Trained entirely on domestic Huawei Ascend chips — hardware-independence proof point完全在国产华为昇腾芯片上训练——硬件自主的证明点
✕ Cons缺点
- 200K context trails the 1M class of 2026 open flagships200K 上下文落后于 2026 年开权重旗舰的 100 万级别
- Already superseded on scores by GLM-5.2/5.3 — compare before locking in分数已被 GLM-5.2/5.3 超过——锁定前请对比
- 744B self-hosting is a multi-GPU project even in FP8744B 自托管即使 FP8 也是多卡工程
Use Cases应用场景
Popular ways people use GLM-5.人们使用 GLM-5 的常见场景：
🏗️ End-to-End Project Construction端到端建项目
From brief to working project across many files and iterations — the agentic-engineering pitch, not snippet autocomplete.从需求到能跑的项目，跨多文件、多轮迭代——这是 Agent 工程的卖点，不是代码片段补全。
🔧 Refactors & Deep Debugging重构与深度排障
Large refactors and deep debugging runs where goal consistency across steps matters more than single-turn fluency.大型重构与深度排障：多步目标一致性比单轮流畅度更重要。
🤖 Multi-Step Tool Workflows多步工具链路
Agent workflows that plan, call tools, check results and iterate — with sparse attention keeping long-context cost down.规划、调工具、验结果、再迭代的 Agent 工作流——稀疏注意力把长上下文成本压下来。
🏢 Self-Hosted Enterprise Coding企业自托管编程
MIT weights inside your own infrastructure for coding assistance without sending code to a vendor API.MIT 权重跑在自己的基础设施里，做编程辅助、代码不送供应商 API。
Similar AI Tools相似 AI 工具
If GLM-5 doesn't fit your needs, here are other popular AI tools you might consider — another open-weight coding flagship at DeepSeek, the closed-model coding reference at Claude, and a high-volume open MoE at MiniMax M3.1 Flash.如果 GLM-5 不符合你的需求，可以看看这些热门 AI 工具——另一开权重编程旗舰见 DeepSeek，闭源编程参照见 Claude，大批量开 MoE 见 MiniMax M3.1 Flash。