GLM-5.3評測：MIT免費開源，Coding跑分暴增6倍 | GLM-5.3 Review: MIT-Licensed, 6x Coding Benchmark Jump

9月 06, 2026

By Kit 小克 | AI Tool Observer | 2026-09-07

🇹🇼 GLM-5.3評測：MIT免費開源，Coding跑分暴增6倍

GLM-5.3 是中國智譜 AI（Zhipu AI／Z.ai）在 2026 年 8 月底釋出的開源權重模型，主打「最強開源程式碼代理」，重點是它完全沿用 GLM-5.2 的基礎模型，這次進步全靠後訓練（post-training）硬幹出來，沒有重新預訓練，卻換來跑分暴增的結果。

GLM-5.3 跑分：Terminal-Bench 進步6倍不是誇張

官方數字乍看很浮誇，但這次有第三方跑分佐證：GLM-5.3 在 Terminal-Bench 3.0 的分數從前代的 4.6 分衝到 28.3 分，等於進步超過 6 倍，在所有開源模型中排名第一。在強調長流程代理任務的 Agents' Last Exam 榜單上，GLM-5.3 同樣是開源模型第一名。官方自己的內部 Code Bench 也宣稱比 GLM-5.2 進步 50%。

要注意的是，Terminal-Bench 這種基準測的是能不能在終端機裡完成多步驟任務，跟一般人熟悉的單題程式碼生成測試（如 HumanEval）不是同一件事，更接近實際跑 agent 幹活的場景，這也是為什麼這次進步幅度看起來特別誇張。

架構：744B 參數 MoE，40B 啟動，200K 上下文

總參數量：約 744B（Hugging Face 上 BF16 版本實測約 753B）
啟動參數：每次推論約 40B（Mixture-of-Experts 架構）
上下文長度：200K tokens
授權條款：MIT License，沒有營收門檻、沒有使用限制，是目前業界最寬鬆的授權之一

另外還有一個小老弟 GLM-5.3-Flash，320B 總參數、18B 啟動，是從全新基礎模型訓練，不是 GLM-5.2 的後訓練版本，主打更輕量的自架部署。

意外附贈的資安能力值得留意

GLM-5.3 在 CyberGym（漏洞挖掘基準）拿下開源模型第一，在滲透測試相關的 exploitation benchmark 上分數是 GLM-5.2 的兩倍以上。官方把這稱為 emergent（湧現）能力，意思是他們並沒有特別針對資安去訓練，這是 coding 能力提升的副產品。這對防守方是好消息（AI 能更快找漏洞），但也代表攻擊門檻在下降，兩面刃是老問題但每次看到跳這麼快還是得提醒一下。

自架建議

744B 參數的完整版對硬體要求不低，個人開發者想跑通常要多卡伺服器等級的顯存；GLM-5.3-Flash 因為只有 320B 總參數、18B 啟動，對自架更友善，是大部分人實際會選的版本。MIT 授權代表可以直接商用微調，不用擔心像某些 open weight 模型那樣藏著營收分潤條款。

好不好用，試了才知道。

🇺🇸 GLM-5.3 Review: MIT-Licensed, 6x Coding Benchmark Jump

GLM-5.3 is the open-weight coding model Zhipu AI (Z.ai) released in late August 2026, and the headline claim is that it is the strongest open-weights coding model available. The interesting part: it reuses the exact same base model as GLM-5.2, all the gains come purely from post-training, no new pretraining run.

GLM-5.3 Benchmarks: A 6x Terminal-Bench Jump

The numbers sound like marketing hype until you check the benchmark: GLM-5.3's Terminal-Bench 3.0 score jumped from 4.6 to 28.3, a 6.2x improvement, ranking first among all open-source models. It also tops Agents' Last Exam, a benchmark focused on long-horizon agentic tasks, among open models. Zhipu's own internal Code Bench claims a 50% improvement over GLM-5.2.

Worth noting: Terminal-Bench measures whether a model can complete multi-step tasks inside an actual terminal, closer to real agentic coding work than single-shot code generation tests like HumanEval. That is part of why the jump looks so dramatic; it is testing a different skill than what most people associate with coding benchmarks.

Architecture: 744B MoE, 40B Active, 200K Context

Total parameters: ~744B (the BF16 repo on Hugging Face reports ~753B)
Active parameters: ~40B per token (Mixture-of-Experts)
Context window: 200K tokens
License: MIT, no revenue threshold, no field-of-use restriction, one of the most permissive licenses in the industry right now

There is also a smaller sibling, GLM-5.3-Flash, at 320B total / 18B active parameters, trained from a fresh base model rather than post-trained from GLM-5.2, aimed at lighter self-hosted deployments.

The Emergent Cyber Capability Worth Watching

GLM-5.3 tops CyberGym, a vulnerability-discovery benchmark, and more than doubles GLM-5.2's score on exploitation-related benchmarks. Zhipu calls this emergent, meaning they did not specifically train for security, it fell out of the coding improvements. That is good news for defenders (faster automated vuln-hunting) but it also means the barrier to offensive tooling keeps dropping. It is the same double-edged pattern we keep seeing, but the size of this particular jump is worth flagging.

Self-Hosting Notes

The full 744B model needs serious multi-GPU hardware, not something most individual developers can casually run. GLM-5.3-Flash, at 320B total / 18B active, is the far more realistic option for self-hosting. The MIT license means you can fine-tune and ship commercially without hunting for hidden revenue-share clauses that some open weight releases still bury in their terms.

好不好用，試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士