Zhipu AI's GLM-4.5: Open-Source LLM Rivals Western Models

Chinese AI powerhouse Zhipu AI has unveiled its latest advancements, the GLM-4.5 and GLM-4.5V model families, signaling a significant stride in open-source large language models designed for complex logical reasoning, sophisticated programming, and autonomous agent tasks. These new models are engineered to tackle a range of practical applications, from generating interactive mini-games and intricate physics simulations to autonomously producing presentation slides with integrated web search capabilities, and even developing complete web applications encompassing both front-end and back-end functionalities.

The multimodal variant, GLM-4.5V, extends these capabilities by incorporating advanced image and video analysis. This version can reconstruct entire websites from mere screenshots and perform screen operations, enabling highly autonomous agent behaviors. Users can explore these features through a ChatGPT-style interface available freely on chat.z.ai after a simple login.

Zhipu AI’s new lineup comprises three distinct models: the robust GLM-4.5, the more resource-efficient GLM-4.5-Air, and the multimodal GLM-4.5V, which builds upon the Air version. Each model offers a dual-mode operational approach, featuring a “think mode” optimized for deep, complex reasoning and a “quick response mode” tailored for rapid, succinct answers.

A key highlight of the GLM-4.5 series is its remarkable parameter efficiency coupled with strong performance. Zhipu AI asserts that GLM-4.5V delivers the most potent capabilities among open-source models of comparable scale. In comprehensive evaluations across twelve distinct benchmarks, GLM-4.5 secured an impressive third place overall among thirteen large language models, and a notable second place specifically for autonomous tasks. Its performance includes a 70.1 percent score on TAU-Bench agent tasks, a 91.0 percent success rate on AIME 24 math problems, and a solid 64.2 percent on SWE-Bench Verified software engineering tasks.

The models demonstrate exceptional efficiency, with GLM-4.5 utilizing just half the parameters of Deepseek-R1 and only a third of Kimi K2, yet consistently matching or surpassing their performance. For web navigation, GLM-4.5 achieved 26.4 percent on BrowseComp, outperforming even the significantly larger Claude Opus 4, which scored 18.8 percent. Even the more compact GLM-4.5-Air model rivals Deepseek R1 in coding tasks, despite its far smaller parameter count.

Underpinning these models is a sophisticated Mixture-of-Experts (MoE) architecture. GLM-4.5 boasts a total of 355 billion parameters, with 32 billion actively engaged at any given time, while GLM-4.5-Air features 106 billion parameters, with 12 billion active. Unlike some contemporaries that favor wider networks, Zhipu AI has opted for deeper architectures with more layers, a design choice based on their research indicating that increased depth significantly enhances reasoning abilities. The models underwent extensive training on approximately 23 trillion tokens, progressing through multiple phases from general data to specialized code and reasoning tasks.

All models are accessible via the Z.ai platform, offering OpenAI-compatible API endpoints. For the developer community, the code is open source on GitHub, and model weights are available for download from Hugging Face and Alibaba’s Modelscope.

Zhipu AI, founded in 2019 by professors from Tsinghua University and headquartered in Beijing, first garnered international attention in 2022 when its GLM-130B model demonstrated performance that surpassed offerings from industry giants like Google and OpenAI. Today, the company employs over 800 individuals, predominantly in research and development. It has attracted substantial investment from prominent Chinese tech firms including Alibaba, Tencent, and Xiaomi, alongside several sovereign wealth funds and international backers like Saudi Aramco’s Prosperity7 Ventures, culminating in a valuation exceeding $5 billion as it prepares for an initial public offering.

However, the rapid ascent of Chinese AI models, including Zhipu AI’s, operates within a unique geopolitical framework. All such models are subject to government censorship, reflecting the priorities and ideological directives of the Chinese administration. This contrasts with the United States, where the government is also exploring restrictions on domestic AI models, albeit driven by a different set of political values. In both instances, these powerful AI systems risk becoming tools in broader culture wars, with distinct ideologies shaping their capabilities and permissible outputs, ultimately leading to similar forms of content control.