Xiaopeng (09868) releases second-generation VLA, introducing new paradigm of "physical world model"

Today, at the "2025 Xiaopeng Technology Day", Xiaopeng Motors (09868) launched the second generation VLA. This model overturns the traditional "visual-language-action" architecture of the VLA model in the industry, eliminating the need for language translation and achieving end-to-end direct output from visual signals to action instructions, bringing a completely new physical model paradigm. The second generation VLA is also Xiaopeng's first mass-produced physical world large model, serving as both an action generation model and a model for understanding and deducing the physical world. It can be cross-domain driven, providing a unified intelligent base for AI cars, humanoid Siasun Robot & Automation, and flying cars. Thanks to the full optimization of the "chip-operator-model" link, Xiaopeng Motors has successfully installed the second-generation VLA with a scale of billions of parameters on the Ultra model with 2,250TOPS, while the industry generally has car-end model parameters in the scale of tens of millions. At the same time, the data training volume for the second generation VLA is close to 100 million clips, equivalent to the total sum of extreme scenarios that a human driver would encounter in 65,000 years. Based on breakthroughs in computing power and model architecture, Xiaopeng's intelligent driving has seen significant evolution. With the second generation VLA, Xiaopeng has launched "Xiaolu NGP", where the average takeover distance on complex roads has increased by 13 times. The industry's first "navigation-free automatic driving assistance" Super LCC+ human-computer co-driving has been introduced, which can be activated globally without relying on navigation, demonstrating the model's generalization learning and emerging intelligence capabilities. He Xiaopeng believes that in the future, large models will become the "operating system of the physical AI world" and the foundation for all physical AI applications. Based on large computing power, large data, and large models, AI will evolve and self-revolutionize at an unimaginable speed. The release of Xiaopeng's second generation VLA large model is an important step for Xiaopeng to move from intelligent driving to embodied intelligence. The Xiaopeng second generation VLA will start pioneer co-creation experience in December 2025, and will be fully deployed with the Ultra model in the first quarter of 2026. At the event, He Xiaopeng announced that Volkswagen has become Xiaopeng's strategic partner for the launch of the second generation VLA. At the same time, Xiaopeng's Turing AI chip has been designated by Volkswagen.