Push Coming Soon | He Xiaopeng Live Detailed Explanation of XPeng's Second-Generation VLA

News Report (Reporter Liu Zhao) — On March 16, just before the second-generation VLA update, XPeng Motors hosted an AskMeAnything live session. XPeng Chairman and CEO He Xiaopeng, and General Intelligence Center Director Liu Xianming attended to answer core user questions, clarify the update schedule, technical capabilities, and future upgrade plans, and remotely verified its real-world road test performance.

The 22 questions addressed during the live session all focused on users’ main concerns about the second-generation VLA. He Xiaopeng shared that since the nationwide launch of test drives at 732 XPeng stores on March 11, market feedback has exceeded expectations, with test drive numbers doubling, Ultra model sales share significantly increasing, and many users reporting that the driving performance now rivals that of professional drivers.

During the live, the company clarified the phased rollout plan for the second-generation VLA: starting March 19, it will be gradually pushed out, initially to P7 Ultra owners, followed by G7 and X9 Ultra, with all P7, G7, and X9 users receiving updates within this month. From April, the update scope will expand to include 2026 P7+, G9, G6, and the Ultra and UltraSE versions of XPeng G7, covering pure electric and super extended-range models. Regarding user concerns about version differences, the company stated that the Ultra version is designed for L4 capabilities supporting full-scenario navigation; the Max version focuses on high-frequency scenarios like highways and urban main roads, offering industry-standard L2 assisted driving. The distilled version is planned for release in the second half of this year. Additionally, on March 18, XPeng will release a single Turing Max version of P7 and a dual Turing UltraSE version.

To address user concerns about mass-produced versions, Liu Xianming clarified that there are no special or limited editions; the mass production version is the 28th branch of the fourth version, which is more stable and consistent than the media test drive version. He Xiaopeng also mentioned that the released version has been thoroughly refined to achieve a score of over 80, utilizing AI large models to reduce reliance on traffic rules and avoid discrepancies between test drive and actual use experiences.

During the live, He Xiaopeng and Liu Xianming remotely connected with media personnel participating in the “5000 km Full-Scenario Intelligent Driving Challenge Across China.” The challenge started in Kashgar and is heading to Shanghai, with over 3,000 km traveled so far, covering highways, national roads, urban streets, and rural roads, enduring harsh weather such as crosswinds, heavy rain, and sandstorms. Participants reported that the second-generation VLA performs steadily in complex scenarios, with single-person, single-day intelligent driving reaching 1,000 km, requiring only occasional takeover, greatly freeing drivers’ attention. Its generalization ability across all scenarios exceeds initial expectations. Liu Xianming stated that the second-generation VLA achieves “nationwide usability upon launch” by abandoning traditional rule-based models, fundamentally restructuring the technical paradigm to enable the system to understand physical world laws and possess cross-dimensional generalization.

In response to trending online videos of the second-generation VLA in real-world tests, the company provided a positive reply. Regarding the video showing the system recognizing a child obstacle and slowing down in advance, Liu Xianming explained that the system completed recognition, deceleration, and rerouting planning but did not brake immediately; future improvements will focus on optimizing long-tail scenarios and enhancing safety in unexpected situations. He Xiaopeng emphasized that ensuring safety and handling various abnormal scenarios are core to achieving L4 autonomous driving. Regarding the video where navigation was not updated for road closures and the system paused for 14 seconds without disengaging, Liu Xianming explained that this demonstrates the system’s ability to infer multiple routes via AI and select feasible solutions, a capability that rule-based models cannot achieve.

On the technical front, Liu Xianming explained that autonomous driving’s core is a physical AI problem. L4 capabilities depend on models, computing power, data, and ontology. XPeng develops all core components in-house, including foundational models, chips, and compilers, optimizing to deliver the impressive performance of the second-generation VLA. Regarding claims of “five times industry-leading,” he said this conclusion is based on internal all-scenario testing, where, on the same route and time, competitors’ systems are used 4-6 times more often than XPeng’s, which directly reflects in higher driver emotional stability, fewer sudden braking events, and smoother driving. He Xiaopeng revealed that the team aims to improve system capabilities by another 5-10 times by the end of the year. The original goal of developing the second-generation VLA was to create a driver-assist system that even older users could trust. The development followed a “solve extreme cases first, then standard cases” approach—initially raising the system’s upper limit, then improving overall experience.

Additionally, the company addressed the VLA’s night and adverse weather performance. The system uses “human-like perception” to simulate human eye and brain processing, with camera night perception surpassing human vision, and clearer imaging in rain and snow. In extreme scenarios, model decision-making remains unaffected. The system also proactively reduces speed on slippery roads and in low-light conditions, combining purely visual solutions with high computational power and large models to improve decision-making efficiency and safety.

Regarding the development of autonomous driving in China and the US, He Xiaopeng and Liu Xianming both stated that China and the US are currently in the top tier globally. China’s complex road scenarios and dense traffic participants provide excellent conditions for refining autonomous driving technology. Coupled with rising AI talent, supportive policies, and large data volumes, Chinese companies are poised to lead in the global physical AI race. Liu Xianming added that Tesla is a peer, not an opponent, and XPeng is willing to share R&D experiences with industry partners to promote collective progress.

For future upgrades of the second-generation VLA, He Xiaopeng said this initial version marks XPeng’s first step toward L4 autonomous driving. Future plans include optimizing experiences in parking garages, campuses, and in-place starts, all powered by AI rather than rules. The AIOS 6.0 version is scheduled for release within this year. XPeng has integrated its cockpit and autonomous driving teams early this year to advance intelligent cockpit, voice interaction, and driving capabilities, with related platform technologies expected to be implemented across multiple models in Q3 and Q4.

He Xiaopeng also stated that 2026 will be the year of global autonomous driving, and the large-scale deployment of the second-generation VLA signifies a new phase in XPeng’s intelligent driving technology. Moving forward, XPeng will use the second-generation VLA as a technological foundation to explore cross-domain integration and continue iterating its capabilities.