XPeng adds "time memory" to its intelligent driving system.

The competition in intelligent assisted driving is now shifting to a battle in complex scenarios.

According to data from the Ministry of Industry and Information Technology, since 2026, the penetration rate of new passenger vehicles equipped with combined driving assistance functions in China has reached 70%, with NOA models exceeding 30%. As pilot-assisted driving gradually becomes more widespread, judgment, response, and stability in complex scenarios are beginning to become key differentiators for automakers.

On August 27th, XPeng Motors released its first major upgrade to the second-generation VLA, with XOS 6.3.0 debuting on the G9L. XPeng also revealed that its Robotaxi, equipped with the second-generation VLA, recently obtained remote testing qualifications for intelligent connected vehicles in Guangzhou, allowing it to conduct driverless road tests on relevant Level 1, 2, and 3 test roads. This signifies that XPeng Motors has entered a crucial stage of road verification for driverless operation.

The key word for XPeng's second-generation VLA upgrade is "time," meaning that the model not only processes the current view of the vehicle, but also attempts to incorporate road changes over a period of time into its decision-making and predict the next actions of traffic participants.

Specifically, the new version incorporates the Infini-VLA long-time-series architecture, which can access road information from the previous 30 seconds; the X-Foresight prediction model is used to extrapolate scene changes within the next 6 seconds; and streaming inference allows for a closer integration of perception, computation, and vehicle actions.

XPeng stated that the new version has increased the number of parameters in the edge model by 3.5 times, improved the response speed by 300%, and enhanced multi-dimensional comprehensive security capabilities by 20 times.

The "understanding time" mentioned here does not mean that AI has acquired an abstract concept of time, but rather that the model can continuously process changes in the environment. For example, when a vehicle encounters a pedestrian turning back, a car in front braking suddenly, or another car cutting in front, recognizing only the current frame is often insufficient; it is also necessary to combine previous motion trajectories to judge the next risk. For mass-produced assisted driving, this is an engineering upgrade with clear application value.

Whether technology can become competitive depends on its integration into vehicles. The G9L began pre-sales on August 11, with a pre-sale price of 259,800 yuan, offering both pure electric and range-extended versions. XPeng also plans to integrate the intelligent driving model and cockpit model in this vehicle to achieve functions such as voice-activated parking and nearby parking, and will apply some of its Robotaxi development achievements to the mass-produced vehicle.

XPeng is attempting to expand its large-scale model from a single driving algorithm to a comprehensive intelligent foundation for the entire vehicle. If driving, parking, cockpit, and body control can share model capabilities, it can reduce information silos between different systems and facilitate continuous upgrades via OTA (Over-The-Air). However, cross-system collaboration also places higher demands on software and hardware compatibility and security verification.

This upgrade didn't start from scratch. The second-generation VLA was already being rolled out to some models in March of this year, and XPeng's X-World model has been used for closed-loop simulation, online reinforcement learning, and model evaluation. Compared to a demonstration at a press conference, data acquisition, simulation verification, and continuous iteration capabilities are the foundation that determine the model's upper limit.

Competition within the industry is also accelerating. Li Auto has already pushed out VLA driving functionality via OTA; Huawei released ADS 5 in April this year, introducing cloud-based multi-agent training, online reinforcement learning, and vehicle-side risk scenarios.

Different companies use different technical names and evaluation criteria, but they are generally aligned in their direction: from rule-driven and localized models to systems driven by data, computing power, and large-scale models.

Continued R&D investment has laid the foundation for XPeng's participation in this competition. In the second quarter of 2026, the company's R&D expenses were RMB 2.91 billion, a year-on-year increase of 32.1%; during the same period, it achieved revenue of RMB 19.74 billion, a year-on-year increase of 8%, and the overall gross profit margin reached 20.7%.

As of the end of July, XPeng's cumulative deliveries had exceeded 1.2 million vehicles. This ever-expanding vehicle fleet helps accumulate more real-world road data, supporting model training, problem identification, and subsequent iterations. If the same technological foundation can be reused across mass-produced vehicles, Robotaxi products, and other products, its R&D investment is expected to achieve even greater economies of scale.

Judging from the current progress, XPeng's second-generation VLA is shifting from single-point function upgrades to model, chip, data, and whole-vehicle collaboration. The G9L will be an important model for the new version to be tested by the market. However, "L4 capability decentralization" is more of a description of the technology source and experience. The mass-produced functions at this stage are still a combination of driving assistance, and the driver needs to continuously pay attention to the road.

In December 2025, He Xiaopeng stated that if XPeng VLA achieved the overall performance of Tesla's FSD V14.2 in Silicon Valley by August 30, 2026, he would set up a Chinese-style canteen in Silicon Valley, modeled after the restaurant at XPeng's headquarters; if this was not achieved, Liu Xianming, the head of XPeng's autonomous driving division, would be "punished."

Rather than how the bet will ultimately be fulfilled, the actual performance of the G9L after being equipped with the new version and user feedback are more worthy of observation.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.