Xpeng Motors has officially launched the XOS 6.3.0 update, introducing its second-generation VLA model. The update transitions the vehicle's perception system from 3D spatial recognition to 4D spatiotemporal understanding, enhancing trajectory prediction and decision-making capabilities.
The rollout includes the VLA Lite version, which now enables the second-generation VLA architecture on models equipped with a single Turing Max chip. This deployment is supported by learning-based token compression and distillation techniques, allowing high-quality visual processing on lower-compute platforms.
A key feature of the new version is the X-Foresight predictive world model, which can anticipate traffic behavior up to six seconds in advance. Additionally, the Infini-VLA long-sequence architecture allows the system to retain context from the previous 30 seconds of driving data.
During an event in August, Xpeng noted that its second-generation VLA is built on a unified technical foundation. This architecture bridges L2 to L4 autonomous driving capabilities, facilitating the company's expansion into higher-level automated driving systems.
