XPeng announces second-generation VLA for Ultra versions of its 2026 P7+, G7, G6 and G9

XPeng said at its 2026 Global New Product Launch Event on January 8, 2026 that the Ultra versions of the 2026 P7+, G7 Super Range Extender, 2026 G6 and 2026 G9 will be equipped with its second-generation VLA, which the company claims is the industry's first physical world large model capable of achieving preliminary L4-level capabilities. He Xiaopeng stated that current "end-to-end small models" might handle 80% of roads well, but the remaining 20% of scenarios would prevent them from ever achieving true autonomous driving. The second-generation VLA eliminates the traditional "language translation layer" and, paired with 2250 TOPS of effective on-vehicle computing power, a 30,000-card cloud computing cluster and nearly 100 million video training data points, enables features such as recognizing gestures to initiate parking, crawling into main roads, identifying irregular vehicles in construction zones, and crossing solid lines to detour from oncoming lanes. XPeng is also training more L4 capabilities, including starting in place, pulling over, recognizing various types of speed bumps, and multiple autonomous driving styles. Robotaxis equipped with the second-generation VLA software have passed third-party testing agency site evaluations and are about to begin public road testing. According to the plan, the second-generation VLA will officially be deployed in vehicles in the first quarter. For the G6 and G9, which come standard with one Turing chip, XPeng offers intelligent upgrade packages: the Ultra SE, with two Turing chips and support for the second-generation VLA, is priced at 12,000 RMB, and the Ultra, with three Turing chips and support for the second-generation VLA and VLM large models, is priced at 20,000 RMB.