The Pulse
Alibaba Targets 10-Trillion-Parameter Qwen Model With New Chip
Alibaba says its Qwen 4.5 and Qwen 5 roadmap could reach 5 trillion to 10 trillion parameters. The company also unveiled the Zhenwu V900 AI processor and plans to expand Alibaba Cloud’s global data-center capacity beyond 20 gigawatts by 203
AI.info Team ·
“The theme of this year’s Apsara Conference is ‘Intelligence Goes Beyond’.”
Joe Tsai, chairman, Alibaba Group
Alibaba is setting out a plan to train models several times larger than the biggest systems it has disclosed, while building its own processor and cloud infrastructure to support them. At its Apsara Conference in Hangzhou on September 22, the company said the future Qwen 4.5 and Qwen 5 model series could scale to between 5 trillion and 10 trillion parameters.
Alibaba said Qwen 4 is already in training. The larger figures belong to the company’s forward roadmap, not to a model available today. The announcement places model development, semiconductor design and data-center expansion inside one programme, giving Alibaba a way to reduce its dependence on outside hardware as it pursues larger training runs and lower-cost inference.
Qwen 4 Starts the Next Training Cycle
The company did not provide a release date for Qwen 4 or say whether the 5-trillion-to-10-trillion range refers to a dense model, a mixture-of-experts system or both. Alibaba described the figures as the projected scale of the Qwen 4.5 and Qwen 5 series, making the announcement a roadmap rather than a product launch.
Alibaba also highlighted work on recursive self-improvement. It said Qwen3.8-Max completed 33 automated improvement cycles over more than a month, covering pipeline design, data validation, experiments and error diagnosis. The updated model’s Artificial Analysis score rose from 40 to 45, according to the company.
In a separate chip-design experiment, Alibaba said a model spent more than 60 hours working through the design lifecycle and made over 10,000 electronic-design-automation tool calls. The resulting chip bus modules reduced area by 42% without a performance reduction, the company said. Alibaba presented the result as evidence that its models can assist with parts of the hardware development process as well as operate on the finished systems.
Zhenwu V900 Brings More Memory to Alibaba’s Stack
T-Head, Alibaba’s chip-design unit, introduced the Zhenwu V900 for AI training and inference. Alibaba says the processor delivers three times the performance of the Zhenwu M890, which was released in May, and includes 216 gigabytes of GPU memory and 1,200 gigabytes per second of inter-chip bandwidth.
The V900 supports FP8 and FP4 data formats alongside higher-precision workloads. Those specifications matter because large models must distribute computation across many accelerators, while inference operators often trade numerical precision for lower cost and faster response times. Alibaba says the chip is intended to serve both demanding training jobs and lower-precision inference.
Mass production and commercial release are scheduled for the first quarter of 2027. The company says its Zhenwu chips already serve more than 650 customers in sectors including automobiles, finance, large language models, embodied intelligence, energy and manufacturing.
Alibaba also showed a supernode server combining the V900 with its ICN Switch, Panmai SmartNIC and Zhenyue SSD controller. The system is designed to support clusters of up to 500,000 cards, according to the company. That figure describes the planned architecture’s scale, not a disclosed production deployment.
20 Gigawatts of Cloud Capacity by 2032
Alibaba’s hardware plans sit inside a broader expansion of Alibaba Cloud. Chief executive Eddie Wu said the company aims to operate more than 20 gigawatts of global data-center capacity by 2032, a target intended to meet rising demand for model training and AI services.
“With this in mind, our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW, fueling the industry’s exponentially rising demand for AI,” Wu said in the company’s announcement.
Alibaba did not attach a spending figure to the new 20-gigawatt target in the release. The plan nevertheless indicates the physical scale required for the Qwen roadmap: larger models need more accelerators, faster connections between them, high-throughput storage and enough electricity and cooling to keep those systems operating at full capacity.
Alibaba Pairs Chips With an Agent Platform
The company’s announcements extend beyond model training. Alibaba Cloud introduced an agentic cloud structure built around model infrastructure, enterprise agent deployment and context services. Its AgentCore platform is designed to help businesses build, operate and monitor AI agents, while Agent Context connects documents, business systems, chat records and multimodal data for real-time use and long-term memory.
Alibaba said its context service can cut token usage by up to 67% in customer service, coding and data-analysis scenarios. It also said its upgraded OpenLake data-lakehouse architecture reduces total costs by 38% and query response times by 40% compared with traditional architectures. Those figures are company claims tied to Alibaba’s own systems; the release does not provide independent benchmark details.
The immediate hardware product is the V900, but the larger strategic bet is the connection between Alibaba’s chips, cloud systems, Qwen models and enterprise applications. Qwen 4 is in training now, while the 5-trillion-to-10-trillion-parameter target belongs to later Qwen generations. The V900 is not scheduled for commercial release until the first quarter of 2027, leaving Alibaba’s most ambitious model plans dependent on infrastructure that is still being prepared.