Li Auto Unveils New Autonomous Driving Model MindVLA-o1; Li Xiang Says Self-Driving Is Just the Starting Point of Physical AI

At the NVIDIA GTC 2026 conference, Li Auto unveiled its next-generation autonomous driving foundation model, MindVLA-o1. CEO Li Xiang stated that the model is built upon the company’s self-developed “native 3D ViT” (Vision Transformer) 3D visual encoder, enabling AI to understand the environment directly in real 3D space for the first time—rather than relying on traditional methods that reconstruct 3D information from 2D video. He explained that the slow progress in autonomous driving stems from AI’s lack of a foundational understanding of the 3D physical world, akin to how human children develop spatial cognition early in life. The new model leverages high-resolution, multi-view vision to jointly process spatial geometry and semantic information during the encoding phase, while integrating LiDAR as a high-precision calibration tool to achieve stable perception beyond 500 meters. Powered by Li Auto’s proprietary Mach chip—which delivers three times the computing power of its predecessor—MindVLA-o1 unifies spatial understanding, reasoning, and driving behavior into a single framework, endowing it with “multimodal reasoning” capabilities that allow it to simulate future scenarios before making decisions. Li emphasized that this foundation model will not only serve autonomous driving but will also extend into robotics and other domains, ultimately becoming a general-purpose agent for interacting with the physical world: “Autonomous driving is just the starting point of physical AI.”