XPENG XNGP

What XNGP is

XNGP is the name XPENG uses for its advanced driver assistance system. The company says it develops the full stack in-house, meaning the software that perceives the road and the driving policy that decides what the car should do next are both built inside the company rather than assembled from outside suppliers.^[1] Systems of this kind sit between the driver and the vehicle's controls: they take in sensor data, build a picture of the surroundings, and issue steering, braking and acceleration commands, while the driver stays responsible for the car.

How widely it is used

Deployment figures say little about whether people actually switch a system on. A more telling measure is monthly active users — the share of people who have the function available and use it in a given month. XPENG reported an urban driving monthly active user penetration rate of 84 percent for XNGP in November 2025.^[1] Urban driving is the harder case for assistance software, because city streets mix vehicles, pedestrians, cyclists and frequent junctions in a way that highways do not.

VLA 2.0

XPENG unveiled VLA 2.0 at its 2025 AI Day on 5 November 2025 and said it was expected to enter mass production in 2026.^[1] The central change is architectural. Earlier driving models typically convert what the cameras see into an intermediate representation — often described in language-like terms — and then translate that into a driving command. XPENG says VLA 2.0 removes the language translation step and generates action commands directly from visual signals.^[2] In other words, the model is trained to go from pixels to steering, braking and acceleration without passing through a verbal description of the scene first.

That approach depends heavily on data and compute. XPENG said VLA 2.0 was trained on nearly 100 million clips of real driving video without data annotation — the usual process in which people label objects and events in each clip by hand.^[2] Behind the training runs sits a cloud cluster of 30,000 cards, on which the company runs a 72-billion-parameter base model that completes a full iteration every five days.^[2] The parameter count describes the size of the model; the five-day cycle describes how quickly a round of training, evaluation and update can be turned around.

Rollout

XPENG said it would invite users in China into a pilot programme for VLA 2.0 in late December 2025,^[1] and that the system would be pushed to all Ultra models in the first quarter of 2026.^[2] Alongside VLA 2.0 the company released a Narrow Road NGP function and said takeover mileage on complex narrow roads rose 13 times.^[2] Takeover mileage is the distance a car travels between moments when a human has to intervene, so a rise in that number is one way of describing how much less often the system hands control back.

Where it fits

XNGP is the layer through which XPENG delivers driver assistance to customers, and VLA 2.0 is presented as its next generation: a shift from pipelines that reason in an intermediate, language-like space to models that map visual input straight onto driving actions. How the system performs on public roads in large numbers of cars — rather than in a pilot programme or on a test fleet — is the question the rollout schedule sets up, with the pilot in China in late December 2025 and the push to Ultra models in the first quarter of 2026.^[1][2] The reported 84 percent urban monthly active user penetration rate for XNGP in November 2025 gives a baseline against which that wider use can be compared.^[1]