Alibaba CEO: Multimodal AI Models Should Combine Perception and Interaction

At the Yunqi Conference on September 22, 2026, Alibaba Group CEO Wu Yongming said that multimodal models integrating understanding and generation represent a key research direction for AI. He stressed that a highly intelligent "brain" alone is not enough; AI must also acquire perception and interaction capabilities, so that people can eventually communicate with AI as naturally as they do with other humans, without relying on complex interfaces.

The direction has already taken concrete shape at the conference. Banma Intelligent showcased its full-modality on-device large model, AutoOmni 2.0-23B-A3B, which adopts a mixture-of-experts (MoE) architecture. In smart cockpit scenarios, the model's routine task capabilities are comparable to a cloud model ten times its parameter count, while its complex task processing reaches 90 percent of such a cloud model's performance. A real-vehicle solution based on an on-device and cloud-integrated agent architecture was also demonstrated, handling vague intent, complex multi-task navigation, vehicle control, multimedia, and cross-domain tasks with low latency, privacy protection, and stability.

Banma Intelligent CTO Si Luo said the 23B on-device model has completed adaptation for mainstream high-compute automotive chips, supports stable concurrent execution of more than eight tasks, and delivers five-to-six times faster inference. He added that on-device AI is opening a new interaction paradigm, moving from cars to multiple end devices toward native intelligence. More than 60 percent of automakers deploying on-device models have chosen Banma's AI solution. The event also featured the Freelander 8, the world's first production vehicle based on Qualcomm's 8397 platform with AutoOmni.