GLM-5重新定義開放權重AI：人工智能分析智能指數新基準

GLM-5 剛剛引起轟動，成為第一個在人工智能分析智能指數上得分 50 的開放權重模型，大大縮小了與專有人工智能巨頭的差距。

The Fifty-Point Leap: Open Weights Takes the Lead

五十點的飛躍：開放式舉重取得領先

In a significant shake-up for the artificial intelligence landscape, Z.ai's latest offering, GLM-5, has officially claimed the top spot among open weights models on the Artificial Analysis Intelligence Index. Scoring an impressive 50, GLM-5 is not just an incremental update; it's the first open weights model to ever achieve this benchmark, marking an 8-point jump from its predecessor, GLM-4.7. This achievement signals a dramatic narrowing of the chasm that once separated open-source innovation from its proprietary counterparts, positioning GLM-5 firmly ahead of other frontier open weights models like Kimi K2.5 and DeepSeek V3.2.

在人工智能領域的一次重大變革中，Z.ai 的最新產品 GLM-5 正式佔據了人工智能分析指數開放權重模型的榜首位置。 GLM-5 獲得了令人印象深刻的 50 分，它不僅僅是一個增量更新；它還是一個增量更新。它是第一個達到這一基準的開放式重量模型，比其前身 GLM-4.7 提高了 8 點。這一成就標誌著曾經將開源創新與其專有同行分開的鴻溝大幅縮小，使 GLM-5 牢牢領先於 Kimi K2.5 和 DeepSeek V3.2 等其他前沿開放權重模型。

Agentic Ambition: Mastering Real-World Tasks

代理野心：掌握現實世界的任務

GLM-5 isn't merely intelligent; it's exceptionally capable. The model boasts the highest Artificial Analysis Agentic Index score among open weights models, hitting 63 and ranking third overall. This prowess is largely driven by its formidable performance on GDPval-AA, a benchmark designed for economically valuable work tasks. From crafting presentations to complex data analysis and even video editing, GLM-5 demonstrates a robust ability to tackle real-world challenges, outperforming many peers and placing it just behind industry titans like Claude Opus 4.6 and GPT-5.2 (xhigh) in practical applications. It's a clear indicator that open weights models are increasingly ready for prime-time enterprise deployment.

GLM-5 不僅具有智能性，而且具有智能性。它非常有能力。該模型在開放權重模型中人工分析代理指數得分最高，達到 63，總體排名第三。這種能力很大程度上得益於其在 GDPval-AA 上的出色表現，GDPval-AA 是專為具有經濟價值的工作任務而設計的基準。從製作演示文稿到復雜的數據分析甚至視頻編輯，GLM-5 展現出了應對現實世界挑戰的強大能力，超越了許多同行，在實際應用中僅次於 Claude Opus 4.6 和 GPT-5.2 (xhigh) 等行業巨頭。這是一個明確的跡象，表明開放權重模型越來越適合黃金時段的企業部署。

Smarter, Sharper, and Less Prone to Piffle

更聰明、更敏銳、更不易胡言亂語

Perhaps one of GLM-5's most compelling advancements lies in its drastically reduced propensity for fabrication. The model shows a remarkable improvement on the AA-Omniscience Index, scoring -1 – a staggering 35-point leap from GLM-4.7's -36. This isn't magic; it's a testament to a 56 percentage point reduction in hallucination rates, achieved by the model's intelligent decision to abstain more frequently when uncertain. In an era where AI reliability is paramount, GLM-5 sets a new standard for trustworthiness among models, offering more accurate and less misleading information. It knows what it doesn't know, which, frankly, is a refreshing trait.

也許 GLM-5 最引人注目的進步之一在於它大大降低了製造的可能性。該模型在 AA-Omniscience Index 上顯示出顯著的改進，得分為 -1，比 GLM-4.7 的 -36 飛躍了 35 分。這不是魔法；這是魔法。這證明了模型在不確定時更頻繁地放棄幻覺的明智決策，使幻覺率降低了 56 個百分點。在人工智能可靠性至關重要的時代，GLM-5 為模型的可信度設立了新標準，提供更準確、更少誤導性的信息。它知道它不知道的東西，坦率地說，這是一個令人耳目一新的特徵。

The Open-Source Advantage: Power, Precision, and Price

開源優勢：功耗、精度和價格

Under the hood, GLM-5 is a beast of engineering, scaling to 744B total and 40B active parameters, and integrating DeepSeek Sparse Attention. While its BF16 precision makes for a hefty ~1.5TB footprint, its MIT License and competitive pricing on third-party APIs like Novita AI and Deep Infra Inc. make cutting-edge AI accessible to a broader audience. Running the Artificial Analysis Intelligence Index with GLM-5 costs a mere ~$547 (based on median third-party pricing), proving that top-tier performance doesn't have to break the bank. With increased pre-training data volume (up to 28.5T tokens), GLM-5 is not just a technological marvel; it's a statement. It underscores a growing trend where open weights models are not just catching up but actively challenging the performance and cost-efficiency of their proprietary counterparts, making the AI frontier a more democratic and exciting space for innovation.

從本質上講，GLM-5 是一個工程巨獸，可擴展到 744B 總參數和 40B 活動參數，並集成 DeepSeek 稀疏注意力。雖然其 BF16 精度造成約 1.5TB 的佔用空間，但其 MIT 許可證以及 Novita AI 和 Deep Infra Inc. 等第三方 API 的有競爭力的價格使更廣泛的受眾可以使用尖端人工智能。使用 GLM-5 運行人工智能分析索引的成本僅為約 547 美元（基於第三方定價中位數），這證明頂級性能並不一定要傾家蕩產。隨著預訓練數據量的增加（高達 28.5T 代幣），GLM-5 不僅僅是一個技術奇蹟；這是一個聲明。它強調了一種日益增長的趨勢，即開放權重模型不僅正在迎頭趕上，而且積極挑戰其專有模型的性能和成本效率，使人工智能前沿成為更加民主和令人興奮的創新空間。

So, for those keeping score, the game has changed. GLM-5 isn't just playing; it's leading the charge, proving that the future of AI might just be more open, more intelligent, and certainly more entertaining to watch unfold. We're all here for the ride.

因此，對於那些記分的人來說，遊戲已經改變了。 GLM-5 不只是用來玩，更是用來玩的。它正在引領潮流，證明人工智能的未來可能會更加開放、更加智能，當然也更加有趣。我們都是來兜風的。