Autonomous Driving Shifts Gears: Physics-Based AI Large Models Take Center Stage, Small Models Relegated to Support Roles

YuanRong Technology CEO Zhou Guang stated that traditional small models have low return on investment in intelligent driving and have hit a development bottleneck. In contrast, physics-informed AI large models—those integrated with real-world physical laws—are becoming a key strategic focus for automakers. This technology breaks away from conventional modular architectures, upgrading intelligent driving systems into a unified cognitive framework that enables full-scenario and cross-domain capability reuse through a single foundational model. Currently, multiple Chinese automakers are accelerating their shift toward physics-informed AI large models: XPeng has announced its transformation into a physics-AI technology company; NIO has established a dedicated large-model division to enhance its intelligent driving features; and Li Auto is focusing on the Vision-Language-Action (VLA) large model direction. Compared with traditional small models—which suffer from fragmented development and heavy reliance on manual tuning—physics-informed AI large models can significantly shorten data closed-loop iteration cycles, improve operational efficiency, enable proactive prediction by redefining scene understanding and decision-making logic, and simplify technical architectures. A representative from YuanRong Technology noted that small models still have applicability in optimizing basic functions for entry-level vehicles with limited computing power and can share capabilities from large models via "model distillation." However, in advanced intelligent driving applications, small models can no longer meet requirements, and their gradual phase-out represents an industry-wide trend. Large models have become the critical pathway for intelligent driving advancement, and companies that transition early stand to gain a first-mover advantage.