Zhipu AI, one of China’s pioneering companies in large language model (LLM) development, has made significant strides in the past two years, according to Zhang Fan, the firm’s Chief Operating Officer. Speaking at BEYOND Expo 2025, Zhang outlined the company’s history, current strategies, and perspectives on the evolving AI landscape both domestically and globally.
Founded in 2019, Zhipu was among the first Chinese firms to develop independent large language model frameworks beyond GPT-based architectures. Zhang, who has a background in algorithm engineering and AI product development and previously worked at Tencent and Sogou, emphasized Zhipu’s early commitment to advancing AI capabilities and commercial applications.
Zhipu AI is widely recognized as one of China’s leading AI model developers – often referred to as one of the country’s top six AI unicorns. In April 2025, the company filed for IPO guidance with the Beijing securities regulator, putting it on the path to potentially becoming the first firm in its segment to pursue a public offering.
Discussing the known directions in AI development, Zhang pointed to a clear framework introduced by OpenAI in 2023, which outlines five levels of AGI capabilities. Zhipu AI aligns with this roadmap, which plots a route from basic dialogue generation to collective intelligence. He noted that a current shared challenge in the industry is converting intelligence into productivity. “We’ve turned computing into intelligence,” he said, “but the next challenge is how to transform intelligence into scalable production.”
Zhang also acknowledged the significance of DeepSeek’s emergence in January 2025, describing its rise as a positive development that has deepened market interest and increased overall industry penetration. “Every player in the ecosystem benefits from this acceleration,” he said.
On the unknowns in the field, Zhang highlighted the constantly shifting boundaries of model capabilities. He cited inference models, which can deliver significantly improved performance on specific tasks without changing the base model, as an example of evolving potential. “We still don’t know where the boundaries lie,” Zhang said.
In terms of Zhipu’s commercial strategy, Zhang stressed that the company remains focused on its core competency: building and optimizing foundational models. The firm has developed a broad model matrix over the past two years and significantly reduced model deployment costs, by as much as 2,000 times, to support enterprise adoption.
However, Zhang acknowledged that offering models alone is not enough. The firm aims to build partnerships with domain experts and industry players rather than developing end-use applications internally. “We believe it takes at least five years to truly understand any industry. Our role is to provide foundational intelligence and combine it with sector-specific know-how to generate practical solutions,” he said.