Z.ai’s flagship GLM-5.2 model, released a day after Anthropic disabled worldwide access to its most advanced models, stunned global users after its benchmark performance narrowly trailed leading closed-source models.
“Our mission is to achieve AGI, so our current focus is on improving our model to reach the upper bound of intelligence. All these resources are helping us,” said Qinkai Zheng, technical lead of the company’s CodeGeeX team.
The accompanying surge in investor interest sent the company’s shares soaring more than 2,000% from their blockbuster Hong Kong debut in January, pushing its market capitalisation above HK$1 trillion (US$128 billion) this week.
“This model is comparable to the top closed models,” Zheng told reporters at its head office in Beijing.
“It’s the first time that an open-source model really delivers very solid coding and agent performance that can compare with the leading proprietary AI companies like Anthropic and OpenAI.”
GLM-5.2 now ranks fourth on Artificial Analysis’ LLM Intelligence Leaderboard and second on Code Arena’s front-end coding leaderboard, while operating at roughly one-sixth of the cost of closed US frontier models.
The sudden “unplugging” of closed US frontier models has triggered concern among global allies such as Canada and France, whose leaders criticised the world’s reliance on US-controlled AI infrastructure at the G7 summit last week.
For the first time, a Chinese open-source AI model has come close to bridging the frontier gap with heavily funded Western AI labs. It has also surpassed US open-source models such as Google’s Gemma and Meta’s Llama series despite restrictions on computing power.
Analysts had previously estimated that China’s leading AI models lagged top US models by four to six months.
Z.ai, also known as Zhipu AI, said earlier this month that it plans a dual listing in Shanghai but did not disclose how much it aims to raise.
Specialising in coding and complex tasks
Specialising in coding and complex long-horizon tasks, the model has 750 billion parameters and a one-million-token context window.
The company said in a blog post that the model was released with inference adaptation for a wide range of domestic chip infrastructure, including Huawei Ascend clusters.
Since February, Z.ai’s GLM-5 series has been adapted to run on domestic semiconductors after the US tightened China’s access to advanced Nvidia chips.
“We are trying our best to improve our infrastructure and make the model more efficient across different types of chips,” Zheng said, without disclosing whether it was trained on domestic or foreign chips.
Despite intense domestic price competition, Z.ai has increased prices for its frontier models several times this year, reflecting strong enterprise adoption in China, where its technology is also widely used in public-sector procurement.
In a post on X, co-founder Tang Jie reaffirmed Z.ai’s commitment to open source and described Anthropic’s abrupt withdrawal of its Fable and Mythos models as “deeply regrettable”.
JP Morgan expects the company’s revenue to surge this year and forecasts it will turn profitable by 2028. However, stock exchange filings show that Z.ai still generates only a fraction of the revenue earned by its US counterparts.
“We are trying to lower costs, but because demand is so high, we may still need to increase prices in the future,” Zheng said. “But we want the model to remain accessible to everyone.”
Zheng added that the company will focus on long-horizon tasks and self-evolving autonomous agent systems in future models.
Its next-generation GLM-5.5 model is expected to be released in August and could mark the next major milestone for China’s frontier AI industry.
($1=7.8397 Hong Kong dollars)
