XPeng said its first VLA model upgrade will raise on-device parameters 3.5 times

XPeng presented the G9L, a five-seat sport utility vehicle, at an event in Guangzhou on August 11 and opened pre-orders at RMB 259,800.

Under the heading it gave to intelligence, XPeng said its second-generation vision-language-action model would receive its first major model upgrade in a forthcoming version 6.3.0, which it said would raise the on-device parameter count by 3.5 times and increase perception sensitivity by 300 percent. The company said the version had the largest model scale of any it had released, responded to decisions faster, predicted trajectories further ahead, held a longer effective time sequence and generalized more strongly across vehicle models and countries, and that the results of distillation had exceeded its expectations.

XPeng said version 6.3.0 would also merge the driving and cockpit models, using the term VLA+VLM driving-cockpit fusion, and would push part of the Level 4 experience from its robotaxi program down into production cars. It listed pulling away from a standstill, pulling over to stop, and roaming a campus to find a parking space. Through the vision-language model, the company said, drivers would also be able to give imprecise navigation instructions and ask by voice to pull over or to park in the nearest space.

On the vehicle itself, XPeng said rear-wheel steer-by-wire was standard across the range, with a minimum turning radius of 5.4 meters. It listed triple-redundant brake-by-wire, dual-redundant drive motors, dual-redundant emergency communications and quadruple redundancy on the door release. The company said the G9L had been tested over three years across 26 countries and regions, covering more than 6.74 million kilometers, and met more than 3,000 regulatory items in 64 countries and regions.

Source: XPeng, August 11, 2026.