Yesterday, XPeng Motors held an AI technology sharing session in Hong Kong, unveiling its newly developed 72-billion-parameter large-scale autonomous driving model – the "XPeng World Base Model." This model, built on a multi-modal architecture, integrates visual understanding, chain reasoning, and action generation capabilities. It aims to be deployed to vehicle-end devices via cloud distillation technology, simultaneously empowering AI robots, flying cars, and other ecosystem products.
According to Li Liyun, XPeng Motors' head of autonomous driving, the base model uses a large language model as its backbone, trained with massive driving data and possessing self-evolution capabilities. Through reinforcement learning, the model will continuously improve its decision-making efficiency, aiming to achieve autonomous driving technology that surpasses human capabilities. To support this research and development, XPeng Motors has accelerated its AI infrastructure deployment since 2023, completing the first 10,000-card intelligent computing cluster in China's automotive industry. This cluster boasts a computing power of 10 EFLOPS, maintaining a utilization rate consistently above 90%, with peak efficiency exceeding 98%.
XPeng Motors calls this end-to-end process, from cloud training to on-vehicle deployment, the "cloud model factory," encompassing pre-training, post-training, model distillation, and on-vehicle adaptation. Currently, this factory achieves an average full-cycle iteration every five days, with video training data increasing from 20 million clips to a target of 200 million clips this year. The R&D team has also developed multi-sized base models. The 72-billion-parameter (72B) model boasts 35 times more parameters than mainstream VLA models, signifying a breakthrough in XPeng's autonomous driving computing power reserves.
XPeng Motors utilizes its "cloud model factory" to close the data loop: after pre-training and reinforcement learning in the cloud, the base model is distilled and compressed into a lightweight version for deployment to the vehicle end. This architecture not only supports intelligent upgrades for existing models but also provides fundamental capabilities for cutting-edge products like AI robots and flying cars.
Officially, XPeng Motors started its AI infrastructure (AI Infra) construction in 2024 and has now formed a complete system covering data acquisition, model training, and scenario implementation. Three key achievements have emerged: verifying the sustained effectiveness of the scaling law in the field of autonomous driving, achieving base model vehicle control on retrofitted computing vehicle-ends, and initiating the training of the 72B parameter model and building a dedicated reinforcement learning framework. In the future, the XPeng World Base Model will be deeply integrated into the AI ecosystem, promoting the co-evolution of intelligent vehicles and robotics.
