XPeng's Gen 2 VLA Launches in September with L4 Autonomy, Debuts on G9L

On August 27, XPeng announced that the new version of its second-generation VLA (Vision-Language-Action) system will be fully rolled out in September, covering all Ultra and Ultra SE models, with the XPeng G9L debuting as the first vehicle to feature it. The system is designed to achieve Level 4 (L4) autonomous driving capabilities, support a broader range of mass-production vehicles, and enable global generalization. Built on a unified technical foundation, XPeng has now achieved seamless capability progression from L2 to L4. The new version introduces the MasterAgent—a centralized "vehicle brain"—to integrate VLA and VLM (Vision-Language Model) functionalities within the cockpit and bring select L4-level Robotaxi experiences to mass-produced vehicles. For the first time, XPeng’s physical world foundation model incorporates the time dimension, expanding AI’s understanding of the world from 3D space to 4D spacetime. The on-device model’s parameter count has increased by 3.5x, supporting effective temporal sequences of up to 30 seconds and enabling scene prediction up to 6 seconds into the future, resulting in a 20-fold improvement in multidimensional safety capabilities. Through learning-based token compression and distillation training, XPeng has deployed its foundation model onto lower-compute platforms, significantly enhancing the performance of Turing VLA 2.0 Lite. The XPeng G9L Max will be the first to feature the distilled version of the second-generation VLA Lite, with all Max variants receiving corresponding capability upgrades. Robotaxis equipped with the second-generation VLA have already obtained Guangzhou’s Intelligent Connected Vehicle remote testing certification, allowing them to conduct driverless road tests—without a safety operator in the driver’s seat—on designated test roads in Guangzhou.