China’s GLM-5 challenges Western AI dominance

China’s Z. ai has propelled its artificial intelligence ambitions into the global spotlight with the launch of GLM-5, a 744-billion-parameter open-source large language model that stands toe-to-toe with leading Western counterparts and underscores Beijing’s drive for technological self-sufficiency. Released on 11 February by the company formerly known as Zhipu AI, GLM-5 combines advanced architecture with broad hardware compatibility and permissive licensing, signalling a shift in the global AI landscape.

The GLM-5 model, built on a mixture-of-experts framework and trained on an expanded dataset of roughly 28.5 trillion tokens, is designed for multifaceted tasks including coding, reasoning and agentic operations. Z. ai reports that the model’s performance on internal benchmarks approaches that of leading proprietary systems such as Anthropic’s Claude Opus 4.5, and on certain tests it outperforms Google’s Gemini 3 Pro. Its permissive MIT licence allows unrestricted commercial use and adaptation, making GLM-5 one of the most accessible frontier AI models available to developers and enterprises worldwide.

A standout feature of GLM-5 is its independence from American semiconductor technology. Trained and deployable on domestic Chinese chips from manufacturers like Huawei Ascend, Moore Threads, Cambricon and Kunlunxin, the model sidesteps reliance on high-end Nvidia GPUs that are constrained under US export controls. Z. ai’s use of the MindSpore AI framework further embeds the model within a Chinese software ecosystem, emphasising self-reliant technological development. This strategy exemplifies a broader push within China’s AI sector to mitigate the impact of external trade restrictions and build robust, locally supported infrastructures.

Market reception underscored the impact of GLM-5’s launch: Z. ai’s shares surged nearly 29 per cent on the Hong Kong Stock Exchange following the announcement, reflecting investor optimism about the company’s competitiveness and growth prospects. Industry analysts note that GLM-5’s arrival comes amid an acceleration of model releases and chip optimisation efforts by several Chinese AI firms, signalling more intense competition on both domestic and international fronts.

Benchmarks published by Z. ai highlight GLM-5’s strengths and limitations. On software engineering evaluations, the model achieved a score of 77.8 per cent on the widely referenced SWE-bench Verified suite, trailing slightly behind Claude Opus 4.5’s 80.9 per cent but outperforming other notable systems in agentic and multitask contexts. The architecture’s efficient handling of long-context sequences—up to 200 000 tokens—positions GLM-5 for applications requiring extensive reasoning over lengthy inputs.

GLM-5’s low cost of deployment also enhances its appeal. Analyses of pricing data suggest inference can cost around US $1 per million tokens, a fraction of the fees charged by some proprietary APIs, potentially lowering barriers for startups and research teams seeking cutting-edge language modelling without prohibitive expenditure. The MIT licence further expands this accessibility by granting broad rights to adapt and deploy the underlying model weights in commercial and research settings.

Despite its technical achievements and open licensing, experts caution against over-interpretation of benchmark results as definitive proof of superiority. Performance under controlled test conditions does not always translate to uniformly strong real-world applications, particularly in generative and reasoning tasks where robustness and safety remain critical hurdles. Some observers also warn that differences between self-reported evaluations and independent assessments could emerge as the model is adopted more widely.

GLM-5’s emergence reflects broader trends within China’s AI ecosystem, where domestic firms and chipmakers are increasingly synchronised in advancing cutting-edge technologies. Industry developments suggest that model and hardware optimisation efforts are happening in near tandem, with architecture adaptations and inference support enabling faster incorporation of new Chinese semiconductor products. These developments sit against a backdrop of regulatory frameworks that increasingly emphasise the alignment of AI systems with government objectives, including content governance and cybersecurity considerations.