Chinese AI Startup Z.ai Releases New Open-Source AI Model

GLM-4.5: Z.ai’s New Open Source AI Model

Chinese AI startup Zhipu AI, now rebranded as Z.ai, recently announced the release of GLM-4.5, an open-source AI model designed for intelligent agent applications [1][2]. In contrast to the logic underlying existing AI models, the company stated that its new model is built on “agentic” AI principles, meaning that the model automatically divides a task into sub-tasks to complete it more accurately [2]. This new model is open-source, making it accessible for developers to download and use [2]. CNBC commented that many Chinese tech firms are developing more advanced AI models that are increasingly cheaper to use, echoing key aspects of DeepSeek’s market-shaking breakthrough [2]. The launch comes as competition intensifies among Chinese tech firms in the global AI development race [2]. During the World AI Conference in Shanghai, Tencent released HunyuanWorld-1.0, an AI model that generates interactive 3D scenes from text or image prompts, while Alibaba has launched Qwen3-Coder, an AI model designed for software development tasks [2].

As of July 2025, China has released 1,509 large-language models, accounting for the largest share globally among the 3,755 models launched worldwide [1]. In late June, OpenAI named Zhipu in a warning about the accelerating progress of Chinese AI efforts [1]. The U.S. government has since added Zhipu to its entity list that bans American companies from doing business with it [1]. The company, established in 2019, is said to be planning an initial public offering (IPO) in Greater China [2]. In this article, we will delve into the key features and capabilities of Z.ai's new open-source AI models.

Key Features

Z.ai’s GLM-4.5 is built on a Mixture of Experts (MoE) architecture, with 355 billion total parameters and 32 billion active parameters [3]. This model is designed for large-scale deployment across reasoning, generation, and multi-agent tasks [5]. Meanwhile, the more streamlined GLM-4.5-Air operates with 106 billion total parameters and 12 billion active parameters [4]. The GLM-4.5 series is distinguished by its comprehensive capabilities, natively integrating reasoning, coding, and agentic abilities in a single model to meet the demands of fast-rising agentic applications [4]. The company said that GLM-4.5 is its first foundation model with native Agent capabilities built directly into the core architecture [4].

The model supports dual thinking modes: a "thinking mode" for complex reasoning and tool usage, and a "non-thinking mode" for quick, direct responses [3][5]. The "thinking mode" is ideal for solving complex reasoning tasks such as mathematics, coding, and logical problem-solving [5]. It takes more time but provides more accurate responses [5]. On the other hand, the “non-thinking mode” is optimized for instant responses, making it suitable for casual or general-purpose use [3][5]. Z.ai’s latest AI models are also capable of multi-step reasoning, function calling, and external tool usage [5]. This means that the model can perform tasks like web browsing, slide creation, and website building—all through natural language commands [5]. It can also design apps, generate code, and even build interactive games [5]. The model works seamlessly with existing coding toolkits like Claude Code and CodeGeex, and takes instructions through simple chat [5].

Performance Benchmarks

Z.ai benchmarked its GLM-4.5 models across 12 industry-standard tests, including MMLU, GSM8K, and HumanEval, to assess their performance across a range of tasks [3]. The flagship GLM-4.5 achieved an average score of 63.2, ranking third overall, second globally, and taking the top spot among all open-source models [3]. Meanwhile, GLM-4.5-Air, a more lightweight version, posted a competitive score of 59.8, establishing itself as the leader among ~100B-parameter models [3].

The models also outperform notable rivals in specific areas, achieving a tool-calling success rate of 90.6%, which surpasses both Claude 3.5 Sonnet and Kimi K2 [3]. GLM-4.5 models also delivered particularly strong results in Chinese-language tasks and coding, achieving consistent state-of-the-art (SOTA) performance across open benchmarks [3] The GLM-4.5 series demonstrates significant advantages in generation speed and pricing, with API calls priced as low as USD 0.11 per million input tokens and USD 0.28 per million output tokens, and a high-speed version achieving generation rates exceeding 100 tokens per seconds [4].

Conclusion

Z.ai's launch of the GLM 4.5 model series highlights a key development in making AI more applicable to real-world scenarios, combining advanced agentic capabilities with robust performance across reasoning, coding, and task automation. With its advanced Mixture of Experts architecture, native agent capabilities, and dual-mode reasoning system, GLM-4.5 is purpose-built for complex, multi-step tasks and real-world applications ranging from coding to interactive content generation. Its top-tier performance in industry benchmarks reinforces its technical credibility and positions it as a serious contender in the competitive AI landscape.

Notes and References

1. Reuters. (2025, July 28). China’s AI Startup Zhipu Releases Open-Source Model GLM-4.5 - Reuters. https://www.reuters.com/technology/chinas-ai-startup-zhipu-releases-open-source-model-glm-45-2025-07-28/
2. Cheng, E. (2025, July 28). China’s Latest AI Model Claims to Be Even Cheaper to Use Than DeepSeek - CNBC. https://www.cnbc.com/2025/07/28/chinas-latest-ai-model-claims-to-be-even-cheaper-to-use-than-deepseek.html
3. Razzaq, A. (2025, July 28). Zhipu AI Just Released GLM-4.5 Series: Redefining Open-Source Agentic AI with Hybrid Reasoning - MarkTechPost. https://www.marktechpost.com/2025/07/28/zhipu-ai-just-released-glm-4-5-series-redefining-open-source-agentic-ai-with-hybrid-reasoning/
4. Z.ai. (2025, July 28). Z.ai Releases GLM-4.5, Setting New Standards for AI Performance and Accessibility While Improving Affordability - PR Newswire. https://www.prnewswire.com/news-releases/zai-releases-glm-4-5--setting-new-standards-for-ai-performance-and-accessibility-while-improving-affordability-302514803.html
5. Dogra, S. (n.d.). GLM-4.5: Is it China’s Best Agentic AI Model Till Date? - Analytics Vidhya. https://www.analyticsvidhya.com/blog/2025/07/glm-4-5-and-glm-4-5-air-launched-by-z-ai/