Zhipu unveils GLM-4.5, intensifies China's open-source AI competition

Zhipu AI Unveils GLM-4.5: A New Front in China’s Open-Source AI Ascent

Chinese AI firm Z.ai, formerly known as Zhipu, has officially launched GLM-4.5, a powerful open-source language model engineered specifically for intelligent agent applications. This release intensifies China’s burgeoning open-source AI race, where domestic startups are rapidly advancing the performance, affordability, and usability of large language models (LLMs).

GLM-4.5 is available in two versions: a flagship model boasting 355 billion parameters (with 32 billion active parameters) and a lighter variant, GLM-4.5-Air, running on 106 billion parameters (with 12 billion active parameters). Z.ai states that GLM-4.5’s performance in reasoning, coding, and agentic capabilities has been rigorously evaluated across 12 representative benchmarks. The company claims GLM-4.5 has secured third place globally and ranks first among both domestic and open-source models based on average scores across these benchmarks. Specifically, GLM-4.5 achieved a 53.9% win rate against top competitors in code generation, and an impressive 80.8% win rate in dedicated coding benchmarks. In real-world agent benchmarks involving API calling, document reading, or search, GLM-4.5 demonstrated a 90.6% success rate, outperforming many closed and open models. It also supports a 128K token context window, enabling it to handle long conversations and multi-document inputs, similar to models like Claude 3 and GPT-4.

A key innovation in GLM-4.5 is its “agent-native” architecture, which integrates core functions such as reasoning, perception, and action directly into the model. This design allows the model to autonomously execute multi-step tasks, generate complex data visualizations, and manage complete workflows. The model also utilizes a Mixture-of-Experts (MoE) framework, activating only a subset of its parameters for each query, which contributes to its computational efficiency while maintaining high performance. Furthermore, GLM-4.5 features dual-mode processing, allowing it to switch between “thinking” and “non-thinking” states for deep, multi-step reasoning or fast, lightweight responses, optimizing for intelligent and contextual output without constant full-throttle compute.

The launch of GLM-4.5 comes as China’s generative AI sector experiences rapid expansion, with the country reportedly developing over 1,500 such models, highlighting its growing ambitions in the global AI landscape. This surge in Chinese open-source AI is seen as a strategic response to geopolitical tensions and efforts to reduce reliance on Western technology.

Z.ai positions GLM-4.5 as a cost-effective alternative to models like DeepSeek’s, which recently gained attention for its disruptive pricing strategies. Z.ai CEO Zhang Peng noted in an interview with CNBC that GLM-4.5 requires only eight Nvidia H20 chips to run, being roughly half the size of DeepSeek’s offering. DeepSeek, for its part, has been praised for its ability to offer services at a fraction of the cost of competitors, sometimes as low as 1/10th, and even operating profitably with substantial margins. This aggressive pricing has pushed other major players like OpenAI to reconsider their own pricing structures, making AI more accessible.

Industry experts view China’s accelerated development of high-performance LLMs as a boon for global enterprises, offering more strategic choices for AI adoption. Neil Shah, vice president for research and partner at Counterpoint Research, highlighted that these models offer comparable or superior accuracy and performance, enabling businesses to tailor AI solutions more precisely to their needs. He also emphasized that GLM-4.5, with its 355 billion parameters, not only excels in Chinese-language tasks but is also open, unrestricted, highly compatible, and can be quantized for broader hardware, facilitating cost-efficient deployments.

The increasing prominence of Chinese models in global AI infrastructure, however, also raises concerns regarding national security, intellectual property, data access, and regulatory compliance amidst ongoing geopolitical tensions, often referred to as an “AI cold war.” Prabhu Ram, VP of the industry research group at Cybermedia Research, acknowledged that China’s aggressive development of models like GLM-4.5 provides cost-effective alternatives to proprietary Western models for enterprises. However, he, along with Shah, stressed the paramount importance of transparency and trust for the global adoption of both Chinese and American models, necessitating stricter and fairer evaluation frameworks for integrating sophisticated AI models into critical enterprise systems.

Zhipu AI’s progress has been notable, with OpenAI itself recognizing Z.ai’s significant strides in securing government contracts in multiple provinces. Reports indicate that Zhipu AI has won 12 public-sector AI contracts worth approximately $307 million USD in 2024, providing AI tools to public-sector clients in countries including Malaysia, Singapore, the UAE, Saudi Arabia, and Kenya, often in partnership with Huawei. Despite being added to the U.S. Commerce Department’s export control entity list in January, limiting its access to American components, Zhipu AI continues to expand its international footprint. China’s broader national strategy, which includes significant state funding and a focus on open-source development, has been instrumental in this rapid growth, with the aim of achieving global leadership in AI by 2030. This open-source approach fosters a collaborative environment, with a growing community of programmers worldwide adapting and improving these Chinese models.