XPeng unveils enhanced vla model with major safety improvements

On August 27, XPeng Group showcased its second-generation VLA large model, which received its first major upgrade with the release of XOS 6.3.0 on the XPeng G9L. The update enables the AI to understand time, transitioning from static 3D space to dynamic 4D spatiotemporal comprehension. The model's edge-side parameter count increased by 3.5 times, surpassing industry standards. XPeng also introduced the Infini-VLA architecture, allowing the AI to remember up to 30 seconds of road events, enhancing decision-making. The new version also features Streaming Inference, boosting response speed by 300%. XPeng's X-Foresight model now predicts events up to 6 seconds ahead, with multi-dimensional safety capabilities improved by 20 times. The Master Agent integrates VLA and VLM in the cockpit, enabling voice-guided parking and autonomous execution. XPeng's Robotaxi recently secured Guangzhou's remote testing approval, while its IRON robot, equipped with three Turing chips, achieved edge deployment. The company also completed a $900 million equity round, valuing its robot business over $6.3 billion. XPeng is deploying the VLA base model on lower-power platforms through distillation, launching it on the G9L Max in September. The company has completed localized testing in Germany, aiming for European regulatory approval by mid-2027.