GLM-5.3-Flash評測：Z.ai開源模型砍到十分之一價格 | GLM-5.3-Flash Review: Open-Source Model at 1/10 the Price

8月 30, 2026

By Kit 小克 | AI Tool Observer | 2026-08-31

🇹🇼 GLM-5.3-Flash評測：Z.ai開源模型砍到十分之一價格

Z.ai(智譜)在8月26日開源了GLM-5.3-Flash，這是一款320B參數的混合專家(MoE)大型語言模型，只啟用180億參數運算，卻擁有100萬token的超長上下文視窗，而且是原生多模態，能吃圖片和影片輸入。最大亮點是價格：輸入每百萬token只要0.15美元、輸出0.5美元，只有前一代GLM-5.3的十分之一，還採用MIT授權可以直接商用。

什麼是 GLM-5.3-Flash？從「隱身」到公開身分

在正式發布前，這個模型其實已經以「Ox Alpha」的匿名身分在OpenRouter上跑了一段時間，因為超長上下文和便宜的速度在社群裡引起討論，直到8月26日Z.ai才揭曉它其實就是GLM-5.3-Flash。這種「先偷跑再公開」的做法，某種程度上是讓模型先接受市場實測，正式上市時已經有口碑墊底。

關鍵技術：混合注意力機制省下大量算力

GLM-5.3-Flash最技術性的賣點，是它首度在開源旗艦模型上把稀疏注意力(sparse attention)和線性注意力(linear attention)混合使用。官方數據顯示，這讓注意力運算量減少約3倍，KV快取縮小到4.4倍以下。白話講就是：處理長文件、長對話時不會像過去那樣又貴又慢。模型本身是在30兆token的多模態語料上重新訓練的基底模型，不是在舊版本上微調而來。

跑分表現如何？跟Claude只差半分

根據Z.ai自己公布的內部程式編碼測試，GLM-5.3-Flash的成績只落後Claude Opus 4.8半分，而且全面超越自家上一代GLM-5.2。當然，廠商自己公布的跑分要打折扣看待，實際體驗還是要自己跑過才算數。

值不值得換？誠實地說

在意長上下文和成本的話，十分之一的價格確實很有吸引力
MIT授權代表可以下載到自己伺服器跑，不用擔心API政策說變就變
目前可在Z.ai自家API(型號名glm-5.3-flash)、GLM Coding Plan、OpenRouter使用，或直接到Hugging Face下載權重
100萬token上下文在真實長文件任務中的「有效」記憶力，常跟跑分公布的理論值有落差，建議拿自己的長文件或程式碼庫先測過

好不好用，試了才知道。

🇺🇸 GLM-5.3-Flash Review: Open-Source Model at 1/10 the Price

GLM-5.3-Flash is the open-source large language model Z.ai (Zhipu AI) released on August 26, a 320B-parameter mixture-of-experts model with only 18B parameters active per token, a 1-million-token context window, and native multimodal input for images and video. The headline number is price: $0.15 per million input tokens and $0.50 per million output tokens, roughly a tenth of its predecessor GLM-5.3, released under the MIT license so you can actually run it yourself.

From Stealth Model to Public Release

Before the official announcement, this model had already been quietly serving traffic on OpenRouter under the anonymous codename Ox Alpha, picking up attention in developer communities for its huge context window and low latency. Z.ai only confirmed on August 26 that Ox Alpha was in fact GLM-5.3-Flash all along — a soft-launch strategy that let the model get real-world stress-testing before the marketing push.

The Technical Bet: Hybrid Attention

The most notable engineering choice in GLM-5.3-Flash is combining sparse attention with linear attention in the same model — a first for an open-weight frontier model. Z.ai claims this cuts attention compute by roughly 3x and shrinks the KV cache by up to 4.4x, which in practice means long documents and long conversations should cost less and run faster than older architectures. The base model was trained fresh on a 30-trillion-token multimodal corpus rather than fine-tuned from an older checkpoint.

Benchmark Numbers: Half a Point Behind Claude

Zs internal coding benchmark puts GLM-5.3-Flash within half a point of Claude Opus 4.8, and ahead of its own previous-generation GLM-5.2 across the board. As always, vendor-published benchmarks deserve some skepticism — real usage is the only test that matters.

Is It Worth Switching?

If long context and cost are your main constraints, the 10x price cut is genuinely compelling
The MIT license means you are not locked into a single API providers policy changes
Available now on Zs own API (as glm-5.3-flash), the GLM Coding Plan, OpenRouter, or as downloadable weights on Hugging Face
A 1-million-token context window rarely delivers the same effective recall on real long documents that benchmark numbers suggest — test it against your own codebase or documents before switching

好不好用，試了才知道 — the only way to know if it works is to actually 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