China · Technology ·

Alibaba outlines AI roadmap: Qwen 4.5 and 5 series planned at 5–10 trillion parameters

Editorial illustration connecting an AI model, accelerator chips and data-center infrastructure in Hangzhou
N環球編輯配圖(非現場影像) · N Globe editorial illustration (not documentary footage) · N環球|N Globe

Alibaba said at its Apsara Conference in Hangzhou that Qwen 4 is in training and the 4.5 and 5 series are projected to scale to 5–10 trillion parameters. It also unveiled the Zhenwu V900 chip and plans more than 20 gigawatts of global data-centre capacity by 2032. Model size, chip performance and production timing are company targets, not independently verified results.

Why it matters

The roadmap links frontier models, domestic AI accelerators and data-centre investment in one infrastructure strategy. Parameter count does not establish real-world capability, and the V900's performance and production targets still require independent testing and delivery evidence.

Full report

Reuters reported at 02:50 UTC on September 22 that Alibaba chief executive Eddie Wu told the Apsara Conference in Hangzhou that Qwen 4 is in training and is expected to have five times as many parameters as Qwen 3-Max. The Qwen 4.5 and Qwen 5 series are projected to scale to 5 trillion to 10 trillion parameters. Those figures are a company development roadmap; they do not mean the models are complete, publicly released or independently benchmarked.

Wu said the larger models are intended to handle more complex, longer-horizon tasks and described artificial superintelligence as a long-term ambition. Parameters are values learned during model training, but a larger parameter count does not by itself prove better reasoning, reliability, safety or operating economics. Alibaba has not released comparable public test results for the planned models.

Alibaba also unveiled the Zhenwu V900 AI chip developed by its T-Head unit. The company says it delivers about three times the performance of the Zhenwu M890 announced in May, calls it China's most powerful AI chip and targets mass production in early 2027. These are Alibaba's claims and schedule. N Globe found no independent benchmark or evidence of volume V900 customer deliveries.

For context, when Alibaba introduced the M890 in May it said 560,000 Zhenwu-series chips had been delivered to more than 400 customers, and its roadmap then placed the V900 in the third quarter of 2027. The new early-2027 production target suggests an accelerated schedule, but manufacturing, supply-chain and validation timelines can change.

Alibaba also plans to raise global data-centre capacity to more than 20 gigawatts by 2032. That target connects model training, inference chips and cloud compute in one infrastructure programme and follows its previously announced three-year, RMB 380 billion AI and cloud investment. Exact construction locations, power sources, capital spending and commissioning milestones were not fully disclosed.

The accompanying image is an N Globe editorial illustration, not footage from the Apsara Conference or a photograph of the Zhenwu V900. Verification status: the company announcement is confirmed; model capability, chip performance and future production results are not independently verified. News value: A-level major news. Origin region: China. Primary section: Technology. Related regions: Asia and World. Source time: September 22, 2026, 02:50 UTC. Report date: September 22, 2026.

READER COMMENTS

Share your view on this story

Be respectful. Maximum 1,000 characters.

Loading comments…