China’s Zhipu AI has debuted its advanced GLM-5.3-Flash AI model, tested extensively on 100,000 locally produced chips, in a move highlighting China’s push for self-reliance in AI hardware and cost-efficient global AI services.
- GLM-5.3-Flash executed entirely on 100,000 Chinese chips
- Model processed 62 trillion tokens during pre-release testing
- Shares rose 12% amid heavy user demand and strong launch
What happened
Zhipu AI successfully launched its latest AI model, GLM-5.3-Flash (formerly Ox Alpha), which operated solely on a cluster of 100,000 domestically manufactured chips. The model underwent a stealth trial where it processed an extraordinary 62 trillion tokens, largely through AI marketplace OpenRouter and the agent platform OpenCode. This set usage records for OpenRouter, boosting the model to the top of global rankings in categories such as coding system performance.
Following the announcement, Zhipu's shares in Hong Kong surged over 12%, reflecting strong investor confidence. The model features 320 billion parameters but activates only a fraction per request to optimize performance and cost. It is also notable for introducing native multimodal processing of visual and textual data in the GLM-5 series.
Why it matters
This launch represents a critical step forward in China's strategic push to reduce dependence on US-based AI hardware providers like Nvidia. By employing a vast network of Chinese-made chips and a specialized inference engine that partitions computing tasks, Zhipu claims to have matched the efficiency and cost-effectiveness of leading Nvidia accelerators, a claim that enhances China's AI hardware credibility in the global market.
The model’s competitive pricing, offered at a fraction of the cost of comparable international models, positions Zhipu to attract global developers, especially amid growing interest in accessible open-weight AI models. The move also intensifies competition within China’s open-source AI ecosystem, with Alibaba releasing comparable multimodal models simultaneously.
What to watch next
Early developer feedback has been mixed, with praise for GLM-5.3-Flash’s coding problem-solving abilities offset by reports of occasional hallucinations, task drops, and slower-than-average generation speeds. Close monitoring of real-world performance and community adoption will be key to validating Zhipu’s efficiency claims and market positioning.
Further developments are expected as Zhipu integrates the model across its platforms and advances multimodal AI capabilities. Meanwhile, competition with Alibaba’s Qwen series and other domestic players may drive rapid innovation and pricing strategies, shaping China’s AI landscape and its role in the international developer community.