XPENG Group recently held a physical AI sharing session and an experience day event for the new version of its second-generation VLA model, centered around the theme of "TIME." During this event, the second-generation VLA large model underwent its first significant upgrade, with the new XOS 6.3.0 version set to make its global debut on the XPENG G9L model. The focal point of this upgrade is empowering AI to comprehend the concept of time, marking a leap from static 3D spatial understanding to dynamic 4D spacetime comprehension. Following the upgrade, the on-device model's parameter count has increased by 3.5 times. It now integrates the Infini-VLA long-sequence architecture, Streaming Inference streaming autoregressive reasoning, X-Foresight predictive world model, and MoT hybrid architecture. This integration has led to a remarkable 20-fold enhancement in multidimensional comprehensive safety capabilities.

In the vehicle's cockpit, the introduction of the Master Agent has achieved seamless integration between driving and cabin functions, bringing some L4-level Robotaxi experiences to mass-produced models. Additionally, the Robotaxi equipped with the second-generation VLA has secured remote testing qualifications in Guangzhou. XPENG's general-purpose humanoid robot, IRON, has also completed on-device deployment. Its robotics business has successfully secured over $900 million in its first round of financing. Moreover, XPENG has employed technical means to deploy the model on low-computing-power platforms, facilitating the global rollout of the second-generation VLA. All these advancements are underpinned by XPENG's comprehensive AI Infra system and data flywheel.