Xpeng Launches 2nd-Gen VLA Model, Volkswagen First Partner

XPeng announced that the second-generation VLA will be open-sourced to global commercial partners to drive industry-wide progress.

On the afternoon of November 5, XPeng officially held the 2025 AI Day under the theme “Emergence.” During the event, the company launched its second-generation VLA model.

The second-generation VLA rewrites the rulebook: it drops the classic “vision-language-action” stack, skips any language translation, and for the first time maps raw visual signals straight to action commands in a single end-to-end pipeline.

According to reports, XPeng’s second-generation VLA was trained using nearly 100 million video clips, covering a vast number of extreme and long-tail scenarios. This dataset is claimed to be equivalent to the total extreme scenarios encountered by an average person driving for 65,000 years. To support this model, XPeng utilized a cloud-based supercomputing cluster including 30,000 cards, with its foundational model reaching 72 billion parameters.

The company stated that the parameter count of the new model has increased more than tenfold compared to the previous generation.

At the execution level, through full-stack optimization of chips, operators, and model layers, XPeng has achieved an average 12x improvement in inference efficiency across three Turing A chips.

Thanks to breakthroughs in computing power and model architecture, XPeng’s intelligent driving capabilities have seen significant advancements in both functionality and user experience.

First, the second-generation VLA introduces the “Narrow Road NGP” feature, which improves the average intervention distance on complex small roads by 13 times, requiring only one intervention every 20 kilometers. XPeng CEO He Xiaopeng shared that in real-world tests on complex small roads, FSD required intervention five times, while the second-generation VLA needed only one.

Additionally, XPeng launched the industry’s first “Navigation-Free Assisted Driving” – Super LCC+ with human-machine co-driving. This function complies with global regulations and can be activated worldwide without relying on navigation. The model can also recognize pedestrian gestures to achieve “wave-to-stop” functionality and understand traffic light logic.

During the event, He Xiaopeng announced the rollout plan for the second-generation VLA: pioneer co-creation testing will begin in late December 2025, with a full rollout scheduled for the first quarter of 2026 alongside Ultra models.

In terms of business cooperation, XPeng has adopted a more open strategy.

XPeng announced that the second-generation VLA will be open-sourced to global commercial partners to drive industry-wide progress. Volkswagen has become the inaugural strategic partner for XPeng’s second-generation VLA, while XPeng’s Turing AI chips have already received nominations from Volkswagen.