Chinese electric vehicle maker Xpeng has begun the full rollout of its second-generation VLA (Vision-Language-Action) system, packaged in the XOS 6.3.0 software update. The company announced the release on September 22, according to Gasgoo, a Chinese automotive news outlet. The update marks a shift from static 3D spatial recognition to dynamic 4D spatiotemporal understanding, as the physical world foundation model now incorporates time into its interpretation framework.
Key Upgrades in XOS 6.3.0
The new version introduces an architecture called Infini-VLA, which retains road environment data from the past 30 seconds. It also features X-Foresight, a predictive world model that forecasts the movement trends of surrounding traffic participants for the next six seconds. The on-device model's parameter count has expanded by 3.5 times, and the end-to-end response speed has improved by 300%.
Alongside the main update, Xpeng has started pushing the second-generation VLA Lite, initially for the Turing Max model. The Lite version is distilled from the high-end model, using learned token compression and distillation training to deploy the foundation model's capabilities on platforms with lower computing power. Xpeng's chairman, He Xiaopeng, revealed that the Lite distilled model has entered mass production and has been tested to rank among the top tier of L2-level assisted driving systems in China.
Functionality and Industry Implications
XOS 6.3.0 also adds new parking features, including head-in and tail-out options, allowing drivers to switch parking direction after selecting a spot, as well as voice-controlled parking exit. The cabin interface now supports split-screen interaction and CarPlay, enabling multiple applications to run simultaneously.
Industry observers note that automakers are exploring ways to reuse the same large-model foundation across vehicles and robots to spread research and development costs, but large-scale commercialization of such technology still requires continuous real-world testing and validation.
