GLM-5.3-Flash評測：神秘模型現形，價格僅Opus的1/30 | GLM-5.3-Flash Review: Mystery Model, 1/30 the Price

9月 03, 2026

By Kit 小克 | AI Tool Observer | 2026-09-04

🇹🇼 GLM-5.3-Flash評測：神秘模型現形，價格僅Opus的1/30

GLM-5.3-Flash 這幾天在開發圈炸出一則有趣的偵探故事：8 月 20 日到 26 日，OpenRouter 上悄悄出現一個代號「Ox Alpha」的匿名模型，免費開放測試。短短六天，它就衝上 OpenRouter 使用量冠軍，還吃下超過 10% 的 OpenCode 流量。開發者靠著比對 tokenizer 輸出、影片 token 消耗模式，以及一些「很 GLM」的錯誤訊息，拼湊出它的真實身分。8 月 26 日，中國 AI 實驗室 Z.ai 正式承認：Ox Alpha 就是他們的新旗艦 GLM-5.3-Flash，權重同步上架 Hugging Face。

GLM-5.3-Flash 規格：320B 參數，MIT 授權開源

GLM-5.3-Flash 是 GLM-5 系列第一個原生多模態模型，支援文字、圖片、影片輸入，採 320B 總參數、18B 啟用參數的 MoE 架構，context window 拉到 100 萬 token。架構上使用混合稀疏與線性注意力機制，搭配 Z.ai 自稱的「Manifold-Constrained Hyper-Connections」，官方數據顯示長文本情境下注意力運算量少了約 3 倍、KV cache 縮小超過 4 倍。最重要的是，這次是完全開源的 MIT 授權，想自架部署沒有任何限制。

跑分逼近 Claude Opus 4.8，價格卻只要三十分之一

在 Z.ai 自家的 Code Bench（max effort）上，GLM-5.3-Flash 拿下 29.0 分，只小輸 Claude Opus 4.8 的 29.5 分；在 DeepSWE v1.1 上，分數從上一代 GLM-5.2 的 46.2 直接跳到 63.4。更狠的是效率：在 High effort 設定下，GLM-5.3-Flash 只用約 5 萬個輸出 token 就拿到 31.4% 的成績，Opus 4.8 卻要燒掉約 12 萬個 token 才有 29.5%。定價方面，輸入 $0.15/M、輸出 $0.50/M、快取輸入 $0.03/M，相較 Opus 4.8 的 $5/$25，輸入便宜約 33 倍、輸出便宜約 50 倍，Z.ai 目前還祭出 5 折上市優惠到 9 月 9 日。

小克怎麼看

先講清楚：上面這些跑分全是 Z.ai 自家公布的，還沒經過大規模第三方驗證，拿來當唯一參考要打折扣。而且公開 benchmark 分數高，不代表在你自己的 codebase、私有工具鏈裡一樣好用——coding agent 最怕的往往是「訓練資料沒看過的專案」。不過「匿名上線六天、免費讓全世界測、六天衝上使用量第一」這招確實聰明，等於讓真實用戶幫忙做壓力測試，而不是關起門來自己刷分。如果你在找便宜、可自架、又想試試多模態 coding agent 的選項，GLM-5.3-Flash 值得排進待測清單。好不好用，試了才知道。

🇺🇸 GLM-5.3-Flash Review: Mystery Model, 1/30 the Price

GLM-5.3-Flash just pulled off one of the more entertaining reveals of the year. Between August 20–26, an anonymous model called "Ox Alpha" quietly appeared on OpenRouter, free to use during the trial window. Within six days it became the single most-used model on the platform and accounted for over 10% of all OpenCode traffic. Developers cracked the mystery by fingerprinting its tokenizer output, matching its video-token consumption to GLM-5V-Turbo, and spotting GLM-flavored error messages. On August 26, Chinese lab Z.ai confirmed it: Ox Alpha was GLM-5.3-Flash all along, and the weights dropped on Hugging Face the same day.

What's Under the Hood: 320B Params, MIT License

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 family, handling text, image, and video input. It's a 320B-total / 18B-active mixture-of-experts model with a 1-million-token context window, built on a hybrid sparse-plus-linear attention architecture Z.ai calls "Manifold-Constrained Hyper-Connections." The company claims roughly 3x lower attention compute and a 4x smaller KV cache at long context versus the base GLM-5.3. It ships under an MIT license — fully self-hostable, no strings attached.

Benchmarks Near Claude Opus 4.8, at 1/30 the Price

On Z.ai's own Code Bench at max effort, GLM-5.3-Flash scores 29.0 versus Claude Opus 4.8's 29.5 — a near-tie. On DeepSWE v1.1, it jumped from GLM-5.2's 46.2 to 63.4. The efficiency gap is the bigger story: at High effort, GLM-5.3-Flash hits 31.4% using roughly 50,000 output tokens, while Opus 4.8 needs about 120,000 tokens to reach 29.5%. Pricing is $0.15/M input, $0.50/M output, $0.03/M cached input — against Opus 4.8's $5/$25, that's roughly 33x cheaper on input and 50x on output. Z.ai is running a 50% launch discount through September 9.

Kit's Take

Worth flagging up front: every benchmark above is self-reported by Z.ai, not yet independently verified at scale — treat the numbers as a starting point, not gospel. Strong scores on public benchmarks also don't guarantee it'll perform the same on your private codebase or internal tooling, which is usually where coding agents actually struggle. That said, the "ship it anonymously, let the internet stress-test it for free, then reveal" approach is a genuinely clever way to get real-world signal instead of just gaming a leaderboard. If you want a cheap, self-hostable, multimodal coding agent to try, GLM-5.3-Flash is worth putting on your list. You won't know until you try it.

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