Alibaba Group CEO Eddie Wu said Tuesday the company’s Qwen team plans to train AI models with between 5 trillion and 10 trillion parameters and unveiled the Zhenwu V900 accelerator, framing a competitive timeline against U.S. frontier labs as Nvidia export curbs limit China’s access to top GPUs.
Speaking at Alibaba Cloud’s Apsara Conference in Hangzhou, Wu said the Zhenwu V900 — from Alibaba’s T-Head unit — delivers about three times the performance of May’s Zhenwu M890 and can scale into clusters of up to roughly 500,000 cards. Mass production and commercial release are targeted for the first quarter of 2027, Wu and company materials said.
On models, Wu said Alibaba aims for the larger parameter range to tackle more complex, longer-horizon tasks. A separate Alibaba statement said Qwen 4 is already in training, with later Qwen 4.5 and Qwen 5 series expected to scale toward 5 trillion to 10 trillion parameters — a figure some coverage has rendered as a range of about 4 trillion to 10 trillion, so treat the top end as company-stated ambition rather than a settled spec. Reuters and Dow Jones have described today’s flagship Qwen 3.8 Max as having about 2.4 trillion parameters; that baseline was not independently verified in Sayer’s research pack, so it is attributed here rather than asserted as established fact.
Alibaba’s Hong Kong-listed shares rose about 5.1% on Tuesday to a roughly one-month high, Technology.org reported. Other market coverage the same day cited a smaller gain near 3%, so the move should be read as early trading reaction, not a settled close.