New model rivals DeepSeek in performance while cutting costs and chip requirements

BEIJING, July 26 — Chinese AI startup Z.ai (formerly Zhipu) has launched its latest artificial intelligence model, GLM-4.5, claiming it is both cheaper to use and more efficient than the widely recognized DeepSeek model.

Z.ai CEO Zhang Peng announced Monday that GLM-4.5 is designed using a breakthrough approach called “agentic AI” — an architecture that allows the model to autonomously deconstruct complex tasks into sub-tasks, enabling more accurate and efficient outcomes.

In a bold move, GLM-4.5 is also fully open-sourced, making it free for developers to download, run, and customize — a strategy that could accelerate adoption both in China and globally.

🧠 Lighter, Smarter, and More Compliant

At half the size of DeepSeek, GLM-4.5 requires only eight Nvidia H20 chips to operate — a significant reduction in hardware demand. The H20 is a China-specific chip designed by Nvidia to comply with U.S. export controls. Although Nvidia recently received permission to resume chip shipments to China, the exact delivery timeline remains uncertain.

Zhang emphasized that Z.ai already has sufficient compute resources and does not plan to acquire additional chips in the short term. However, the startup has not disclosed training costs, saying further details will be shared later.

🚀 China’s Open-Source Push

GLM-4.5 represents another step in China’s push to democratize AI development while circumventing U.S. tech restrictions. By offering a cost-effective, lightweight model that rivals global leaders in functionality, Z.ai aims to capture developer mindshare and compete in the increasingly fragmented global AI market.

The company’s innovation in agentic architecture and commitment to open-source accessibility places it at the center of China’s AI race, especially as enterprises seek alternatives to U.S.-based platforms.