Pushes second-generation VLA, XPeng bets on L4 with both mass-produced cars and Robotaxi simultaneously.

```

Start of the new year, XPeng Motors Chairman and CEO He Xiaopeng stated in the company’s work resumption letter that the DeepSeek moment for autonomous driving has arrived.

Focusing on this core business priority, XPeng’s advancement speed is rapid.On March 2, He Xiaopeng officially announced: XPeng’s second-generation VLA (Vision-Language-Action model) will begin full rollout in the latter part of this month.

This isn’t just a routine software upgrade. In He Xiaopeng’s view, current industry L2 assisted driving is a patchwork of various technical solutions, end-to-end small models have hit a capability ceiling, and intelligent driving R&D has reached a critical watershed moment.

“XPeng’s second-generation VLA is the first version aimed at fully autonomous driving, and it will iterate at an unprecedented speed,” He Xiaopeng predicts. “Fully autonomous driving will arrive completely within the next 1-3 years, and autonomous driving will truly become people’s daily travel habit.”

At the same time, He Xiaopeng also announced that Robotaxi equipped with the model has started public road tests, trial operations will begin within the year, and global delivery will start in 2027.

Face to the future L4 finale, XPeng has pulled the trigger.

Intelligent Driving Steps Out of the “Geek Circle”

“L2 is just a patchwork monster.” He Xiaopeng said bluntly.

In his view, the mainstream assisted driving systems in the industry are essentially splices of different technical solutions: one logic for highway scenarios, another for urban scenarios, parking has another set of rules. This modular stacking may perform well in single scenarios, but once faced with the complexities of the real world, the system’s disjointedness becomes obvious.

XPeng’s second-generation VLA tries to fundamentally answer one question: If the target is L4, why patch along the L2 path?

The answer is a thorough paradigm reconstruction. The second-generation VLA no longer relies on a language model as an intermediate translation layer, but directly achieves end-to-end mapping from visual input to action output.

This means the system can, like a human driver, directly deduce the optimal driving strategy based on visual information, rather than performing serial computations across perception, prediction, planning, and other modules.

This technological route shift is particularly evident at the data layer.

According to Liu Xianming, head of XPeng’s General Intelligence Center, the second-generation VLA has accumulated 50PB of training data, processes about 5.3 billion bytes of visual information per second, and since Technology Day 2025, has completed 468 iterations of model versions.

More noteworthy, the token consumption of the in-car model inference is about 80 times the daily call volume of AI tokens nationwide. Behind these data is XPeng’s belief in the scale law: L4 capability = Model × Computing Power × Data × Entity.

On the computing power level, XPeng’s fully self-developed Turing chip begins to play a key role. Official data shows its computing power utilization rate reaches 82.5%, model inference takes only 80 milliseconds, and the effective computing power of one Turing chip is about equal to 10 Orin-X chips. With joint optimization across chip-compiler-model, compilation efficiency of the base model has increased by 12 times.

Thisis alsoLiu Xianming’s first public appearance since the merger of XPeng’s autonomous driving and smart cockpit centers to form the General Intelligence Center.

He admitted, the R&D process in the past period has been “very painful,”“all R&D in autonomous driving had to be restarted from scratch, going back to the beginning, reframing the autonomous driving problem as a physical AI problem.”

This statement to some extent confirms the depth of this technical upgrade: it’s not an optimization of the existing architecture, but a redo from the underlying logic.

Further, XPeng’s second-generation VLA achieves comprehensive upgrades on three experience dimensions: safe and silky smooth, all-scenario capability, high efficiency.

For smooth safety, the second-generation VLA can recognize various odd-shaped vehicles, detour accident sites, slow down in advance on bumpy roads, yield to small animals at night, and more.

Actual test data show heavy braking decreases 99%, sudden acceleration decreases 98%, safety interventions decrease 60%, road obstacle recognition improves 124%, rear-side vehicle recognition improves 118%, nighttime decision accuracy improves 96%, and comfort of avoidance improves 95%.

All-scenario capability means the second-generation VLA covers campus paths, rural dirt roads, and roads without navigation, can handle narrow passages, rural pothole avoidance, and complex scenarios, and supports P-gear stationary starts for full-journey assistance.

On efficiency, the second-generation VLA improves comprehensive driving efficiency by 23% while ensuring safety and stability. In Guangzhou urban area tests, it completed trips 1 minute faster than the navigation’s 44-minute estimate, at 43 minutes.

For XPeng’s second-generation VLA performance, analysts have given some positive reviews. Morgan Stanley’s analysts believe, “XPeng’s second-generation VLA is a bold leap,” “Tesla will face more Chinese competitors like XPeng who are capable of competing globally in autonomous driving technology.”

Robotaxi Ambition

Outside passenger car intelligent driving, the second-generation VLA’s other application is Robotaxi.

He Xiaopeng revealed that Robotaxi equipped with the second-generation VLA has begun public road tests, will start trial operation within the year, and plans to begin global delivery in 2027. Volkswagen will be the launch customer for this model.

This timetable coincides with a sensitive moment in the Robotaxi industry. 2026 is widely seen as the year of large-scale Robotaxi rollout, with players like Tesla, Baidu, Pony.ai, and WeRide all ramping up deployment.

Tesla plans to mass-produce the CyberCab in Q1 2026;as of February,Baidu’s Robotaxi “Luobo Kuaipao” is operating in 26 cities; Pony.ai’s joint venture with Toyota China plans to deploy thousands of Robotaxi units in 2026.

However, rising industry enthusiasm is matched by public concerns about safety.

Light Chaser CEO Yu Qian recently stated, “China’s Robotaxi may be slower than the US, because it involves labor costs, laws, and multiple other factors.”

Against this background, XPeng’s Robotaxi strategy shows a certain cautious aggressiveness. Cautious in that it chooses to first verify the technology on passenger cars, accumulating data and scenarios from large-scale user adoption; aggressive in that He Xiaopeng has set the timeline for fully autonomous driving at “the next 1-3 years.”

This judgment is based on a rethinking of technical routes. He Xiaopeng believes the industry will leap from L2 directly to L4, with L3 just a transition. He also proposed accelerating regulatory and management policy for the leap from L2 to L4 at this year’s Two Sessions.

If this judgment holds, then current investment in the second-generation VLA will determine XPeng’s position in the next competitive cycle.

Another noteworthy signal is that XPeng’s second-generation VLA plans global delivery in 2027. This means not only technological capability output, but also adaptation to overseas market regulations and road scenarios.

He Xiaopeng has also declared August as the big test for the autonomous driving team: XPeng aims to achieve Tesla’s results in Silicon Valley, but in China.

This benchmarking is no coincidence. The rollout of Tesla’s FSD V12 version early in 2024 greatly encouraged He Xiaopeng, who actively sent key team members to the US for firsthand experience. Smoothness, human-like behavior, thinking ability—these evaluations of FSD now become the goals XPeng’s second-generation VLA pursues.

From a broader perspective, the push of XPeng’s second-generation VLA signals the entry of China’s intelligent driving industry into a new competitive phase.

An industry insider told Wallstreet News: as L2 assisted driving becomes standard and price wars compress profit margins, technological breakthroughs become a necessary choice for top players.

Physical AI, L4 autonomous driving, Robotaxi—once distant concepts—are now being translated into concrete product plans and delivery timetables, paving the way for XPeng to become a top player.

He Xiaopeng’s hope for the future: “After regulations open up, full autonomous driving will let everyone get home safely even after drinking at night, let family members go out without driving themselves, and let cars proactively serve their owners.”

This is undoubtedly a highly attractive vision of the future. But on the road to this vision, technical validation, regulatory improvement, user acceptance, and the shadow of safety incidents are all hurdles that must be overcome. Whether the second-generation VLA can truly usher in the DeepSeek moment for autonomous driving—the answer lies in real user feedback to come.

Risk warnings and disclaimersThe market has risks, investment should be cautious. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment objectives, financial situation or needs. Users should consider whether any opinions, viewpoints or conclusions in this article are suitable for their own circumstances. Those who invest based on this, bear responsibility for themselves. ```