At the 2026 Apsara Conference held this morning, Alibaba Group Chief Executive Officer Eddie Wu shared insights into the company's forward-looking artificial intelligence strategy, revealing plans to expand future AI models to a scale of 5-10 trillion parameters.

Wu noted that over the past year, the technical pathway toward Artificial Superintelligence (ASI) has become increasingly clear. This trajectory, known as Recursive Self-Improvement (RSI), enables models to identify their own shortcomings from real-world feedback, design their own experiments, construct their own data, and evaluate results, thereby driving a continuous cycle of self-evolution. Currently, the Qwen team is actively exploring RSI and has achieved notable progress in this area.

The Qwen team is also advancing research in model architecture and data optimization, with plans to train a brand-new model featuring 5-10 trillion parameters. The objective is to accomplish more complex, long-horizon tasks and take a significant step toward ASI.

However, Wu emphasized that a highly intelligent "brain" alone is insufficient. Artificial intelligence must also master perception and interaction capabilities, understanding sound, images, expressions, and movements much like humans do, while aligning with human culture, aesthetics, and values. Consequently, the development of unified multimodal models that integrate both comprehension and generation has become another key research focus for Alibaba. When AI can communicate like humans, users will no longer need to navigate complex interactive interfaces.

Intelligence is also making its way to end-user devices. On the desktop front, the open-source Qwen2.7B model runs smoothly and supports local deployment for developers and enterprises. On the mobile side, Alibaba has launched Qwen Intelligence, offering AI solutions to partners that enable smartphones to handle a wide range of complex tasks.

Wu further elaborated that Alibaba's T-Head chip division has executed a comprehensive strategy for data center chips, covering the Zhenwu series of GPUs, the Yitian series of CPUs, the Panmai series of smart network interface cards, and ICN interconnect chips. The core semiconductor components required to build ultra-large-scale AI clusters are now fully covered.

At this year's Apsara Conference, T-Head unveiled the domestic AI chip Zhenwu V900, which delivers three times the computing power of its predecessor, the Zhenwu M890. A single cluster built on this chip can scale to accommodate up to 500,000 cards, supporting the training and inference of cutting-edge models. Alibaba stated that due to the maturity of T-Head's chip product line and its widespread adoption by customers, annual shipment volumes are expected to increase significantly.

According to the company, Alibaba is currently conducting joint optimization across chips, servers, supernodes, networks, models, and inference systems to enhance the token production capacity and overall efficiency of its AI clusters. The proprietary M890 AI supernode has already supported efficient inference for large models exceeding 2 trillion parameters and has been deployed at scale on Alibaba Cloud. In the fourth quarter of this year, Alibaba Cloud will add additional service nodes to further expand the supply scale of these supernodes.