China's Zhipu AI Launches GLM-5, Intensifying Race With DeepSeek

Zhipu Fires the Starting Gun

Chinese artificial intelligence developer Zhipu has announced its most ambitious model yet, GLM-5, a seven hundred and forty-five billion parameter large language model built on a Mixture of Experts architecture with forty-four billion active parameters per inference. The model more than doubles the scale of its predecessor, GLM-4.5, which had three hundred and fifty-five billion total parameters.

Built Entirely on Chinese Hardware

In a significant milestone for China's push toward technological self-reliance, GLM-5 was trained entirely on Huawei Ascend chips using the MindSpore framework, achieving full independence from American-manufactured semiconductor hardware. The model was reportedly trained on a massive one hundred thousand chip Huawei Ascend cluster, representing an entirely domestic hardware pipeline amid ongoing United States technology export restrictions.

Competing With the Best in the West

Zhipu says GLM-5 has been measured directly against Anthropic's Claude Opus series, with enhancements in creative writing, programming, reasoning, and intelligent agent capabilities. The model first appeared anonymously on the global model service platform OpenRouter under the codename Pony Alpha, where community testing confirmed its identity through system prompts and tokeniser responses. Early benchmarks suggest performance comparable to or exceeding leading Western models on coding and agentic tasks.

A Crowded Battlefield

The release arrives amid a flurry of model launches from Chinese technology giants. Alibaba has released Qwen3-Max-Thinking, Moonshot AI unveiled Kimi K2.5, and Baidu launched Ernie 5.0. ByteDance is preparing Doubao 2.0 during the holiday period, while DeepSeek is expected to release its V4 model around the seventeenth of February. Zhipu's shares have surged dramatically since its Hong Kong initial public offering in early January, with market capitalisation approaching nineteen billion dollars, nearly triple its listing valuation.

The Bigger Picture

Zhipu founder Tang Jie has acknowledged the challenges ahead, noting that while the company excels in certain areas, disparities remain compared to American AI developers. Nevertheless, the pace of innovation from Chinese labs continues to accelerate, with some analysts arguing that United States chip export controls have paradoxically spurred greater efficiency in Chinese AI training methods.