Beijing Auto Show: Smart Driving Race Shifts to Large Models and Local Implementation

At this year’s Beijing Auto Show, the focus of competition in the intelligent driving sector has shifted from hardware specifications to capabilities in large AI models and scalable deployment. The previous “arms race,” centered on hardware metrics such as the number of LiDAR units and chip computing power, has noticeably cooled. Automakers are now prioritizing in-house large models, data-driven systems, and localized adaptation. Domestic automakers are broadly advancing full-stack in-house development. NIO unveiled its “World Model” (NWM) for autonomous driving, emphasizing a balance between technological iteration and cost control through self-developed solutions. XPeng disclosed that its second-generation VLA intelligent driving system achieves end-to-end integration across perception, decision-making, and planning/control, reducing driver takeover rates in extreme scenarios by 62% compared to the previous generation. XPeng also noted that advanced intelligent driving has become a key factor in vehicle purchase decisions, with Ultra trim users activating intelligent driving features at an average daily rate of 98.52% during their first week of ownership. Joint-venture brands are accelerating efforts to localize their offerings. BMW globally debuted the long-wheelbase version of its next-generation iX3, equipped with an all-scenario navigation-assist system co-developed with Momenta. Mercedes-Benz introduced urban and highway navigation-assist functions powered by a reinforcement learning-based large model in its new S-Class sedan—developed primarily by its China team. Audi’s all-new A6L became the brand’s first mass-produced vehicle in China to feature Huawei’s Qiankun full-stack intelligent driving solution. Volkswagen also launched a new intelligent driving platform based on a localized large model, tailored specifically for China’s complex road conditions. Intelligent driving solution providers simultaneously showcased differentiated technical approaches. Momenta employs a “data flywheel + dual-track synergy” strategy, integrating data loops from both ADAS and Robotaxi operations. Zhuoyu Technology unveiled a native multimodal foundation model capable of independently handling the entire pipeline from perception to decision-making. QCraft proposed a “world model + reinforcement learning” architecture and shifted its strategic focus toward “general-purpose physical AI.” Industry consensus holds that intelligent driving competition has entered a comprehensive phase centered on data-loop capabilities, large-model training efficiency, and engineering implementation maturity. In the future, advanced intelligent driving will rapidly expand into mainstream vehicles priced between RMB 100,000 and 200,000, while facing critical challenges related to model interpretability, data security, and regulatory compliance.