GLM-5.2評測：開源模型解題率99.2%但暗藏跑分陷阱 | GLM-5.2 Review: Open Model Hits 99.2% AIME, Trap Looms

8月 17, 2026

By Kit 小克 | AI Tool Observer | 2026-08-17

🇹🇼 GLM-5.2評測：開源模型解題率99.2%但暗藏跑分陷阱

中國AI新創Z.ai在6月推出的GLM-5.2，最近因為在AIME 2026數學測驗拿下99.2%解題率登上排行榜第一而爆紅——這數字比多數旗艦付費模型都高，而且GLM-5.2是完全開源、MIT授權的模型，任何人都能下載權重自己架設。乍看之下像是「開源模型終於追上閉源」的里程碑，但拆開細看，故事沒那麼單純。

GLM-5.2是什麼：744B參數、只啟用40B

GLM-5.2採用混合專家（MoE）架構，總參數744B，實際推理只啟用約40B，這也是它能把成本壓低的關鍵。規格重點：

授權：MIT，可商用、可離線部署、無地區限制
Context window：最高100萬token
API價格：官方每百萬token輸入1.4美元、輸出4.4美元，第三方平台如OpenRouter更低到0.95/3.0美元
下載：Hugging Face直接開放權重，也有20多個第三方coding工具已整合

根據VentureBeat報導，GLM-5.2在多項長流程coding測試贏過GPT-5.5，價格卻只要六分之一，這才是它真正吸引開發者的地方，而不只是AIME那個99.2%的數字。

99.2%的AIME成績，到底代表什麼

問題出在AIME這類數學競賽題目，網路上流傳大量解題過程與討論串，模型很可能是「背過類似題型」而不是真的展現推理能力。有分析者指出，近期幾個頂尖模型都出現一個共同現象：為了衝高AIME、SWE-bench這類推理型跑分，廠商犧牲了模型的一般知識廣度——GLM-5.2的SimpleQA真實世界知識測驗表現，就明顯不如它的數學分數亮眼。換句話說，GLM-5.2很會解競賽數學題，但你拿它查冷門知識時，未必比舊模型準。

值不值得換：老實講的結論

如果你的場景是程式碼生成、agent長流程任務，GLM-5.2的coding跑分＋開源＋低價格組合，確實值得放進候選名單，尤其是需要自己架設、避開API依賴的團隊。但如果你期待的是「開源模型全面打平GPT級智能」，那99.2%這個數字水分不小，別被單一跑分沖昏頭。

好不好用，試了才知道。

🇺🇸 GLM-5.2 Review: Open Model Hits 99.2% AIME, Trap Looms

Z.ai's GLM-5.2, released in June, is making headlines for hitting 99.2% on the AIME 2026 math benchmark — a score that puts it ahead of most paid flagship models. What makes it notable isn't just the number: GLM-5.2 is a fully open-weight model under an MIT license, meaning anyone can download and self-host it. That sounds like the moment open source finally caught up to closed frontier labs. The real story is more nuanced.

What GLM-5.2 Actually Is: 744B Total, 40B Active

GLM-5.2 is a mixture-of-experts model with 744 billion total parameters but only about 40 billion active per inference, which is how Z.ai keeps costs down. Key specs:

License: MIT — commercial use, self-hosting, no regional restrictions
Context window: up to 1 million tokens
API pricing: $1.40 per million input tokens / $4.40 output from Z.ai directly; third-party hosts like OpenRouter go as low as $0.95/$3.00
Availability: open weights on Hugging Face, plus integration in 20+ third-party coding tools

According to VentureBeat, GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks at roughly one-sixth the cost — arguably a more useful headline than the AIME score.

What the 99.2% AIME Score Actually Means

AIME-style competition math problems circulate widely online with full solutions, so high scores can reflect pattern-matching against memorized solutions rather than genuine reasoning. Analysts tracking this trend note that several frontier models — GLM-5.2 included — appear to be trading general factual knowledge for reasoning-benchmark gains: GLM-5.2s score on SimpleQA, a real-world factual-knowledge test, lags well behind its math performance. In practice, GLM-5.2 is very good at competition-style math, but that does not guarantee it is more reliable than older models for general knowledge lookups.

Is It Worth Switching? The Honest Take

If your workload is code generation or long-horizon agent tasks, GLM-5.2s combination of strong coding benchmarks, open weights, and low pricing earns it a spot on your shortlist — especially for teams that want to self-host and avoid API lock-in, as detailed in this benchmark breakdown. But if you are expecting open source to finally match GPT-class intelligence across the board, the 99.2% AIME number is doing a lot of marketing work it cannot fully back up — see Artificial Analysis independent index for the fuller picture.

好不好用，試了才知道。 (You wont know if it is good 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