Zhipu launches GLM-4.5 model as China ramps up open-source AI race

Chinese AI firm Z.ai, formerly known as Zhipu, has launched GLM-4.5, an open-source language model designed for intelligent agent applications, as competition heats up among China’s fast-growing generative AI startups.

GLM-4.5 will be available in two versions, with the flagship model featuring 355 billion parameters and a lighter variant, GLM-4.5-Air, running on 106 billion parameters.

“Its performance in reasoning, coding, and agentic capabilities has been evaluated on 12 representative benchmarks,” the company said in a statement. “Based on the average score across all these benchmarks, GLM-4.5 has secured third place globally and the first place among both domestic and open-source models.”

The company said GLM-4.5 is its first foundation model with an “agent-native” architecture, meaning core functions such as reasoning, perception, and action are built directly into the model. This design allows it to autonomously execute multi-step tasks, create complex data visualizations, and manage complete workflows.

In June, OpenAI noted that Z.ai had made significant strides in winning government contracts in multiple provinces, underscoring China’s accelerating push to assert itself in the global AI race.

Open-source and low-cost

China’s generative AI sector is expanding quickly, as local startups intensify efforts to improve the performance, affordability, and usability of their open-source large language models (LLMs).

The country has reportedly developed more than 1,500 such models, underscoring its growing ambitions in the global AI race.

GLM-4.5 is positioned as a lower-cost alternative to DeepSeek’s model, which recently drew global attention for its disruptive pricing.

In an interview with CNBC, Z.ai CEO Zhang Peng said that GLM-4.5 requires just eight Nvidia H20 chips to run, noting that the model is roughly half the size of DeepSeek’s offering.

“China’s accelerated development of high-performance LLMs is broadening the strategic choices for global enterprises,” said Neil Shah, vice president for research and partner at Counterpoint Research. “These models, boasting comparable or superior accuracy and performance, enable businesses to tailor AI adoption more precisely to their target applications, regional strengths, and operational focus.”

China’s growing presence in open-source AI is also strengthening its global position, driving innovation and reducing development costs, which is creating new opportunities for enterprise adoption.

“GLM-4.5, for instance, 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, enabling high-performance and cost-efficient deployments,” Shah said.

Impact on enterprises

GLM-4.5 is adding a new dimension to enterprise AI strategy, offering organizations an alternative to proprietary models from the US amid growing interest in open-source tools.

“China’s aggressive development of these models, such as GLM-4.5, is reshaping AI adoption for US and global enterprises,” said Prabhu Ram, VP of the industry research group at Cybermedia Research. “This push is enabling cost-effective alternatives to proprietary Western models.”

For enterprises, this could broaden access to customizable, high-performance AI tools suited for tasks like multilingual processing and agent-driven workflows, especially in cost-sensitive environments.

However, the growing influence of Chinese models in global AI infrastructure comes with concerns.

In the context of ongoing geopolitical tensions and what some analysts refer to as an “AI cold war,” enterprise adoption of non-Western models raises questions about national security, intellectual property exposure, access to sensitive data, and regulatory compliance.

“Consequently, transparency and trust will become paramount for the global adoption of both Chinese and American models across enterprises,” Shah added. “This will necessitate the development of stricter and fairer evaluation frameworks to seamlessly integrate sophisticated AI models into critical enterprise systems and the AI stack.”

Related reading:

• It might be time for IT to consider AI models that don’t steal
• OpenAI’s Skynet moment: Models defy human commands, actively resist orders to shut down
• Alibaba rolls out new AI coding model Qwen3-Coder, says it’s their most powerful
  

>