Guangzhou-based carmaker XPeng said on Monday it is developing a 72 billion parameter large-scale autonomous driving model, named XPeng World Base Model.
At an AI tech conference held on Monday in Hong Kong, XPeng’s autonomous driving head, Li Linyun, explained that the base model is built on a large language model framework, trained with vast amounts of high-quality driving data.
It is a multimodal model with visual understanding, chain reasoning, and action generation capabilities.
Li said at the event that the problem of autonomous driving is essentially a problem of “complex physical AI”, and AI car is the first step of embodied AI.
The executive added that autonomous driving needs to make actions, including dynamic response like acceleration & deceleration, throttle brake, steering wheel and even more human-vehicle interaction with the outside world in the future.
XPeng’s founder and CEO He Xiaopeng had previously stated in March that he believed the company would achieve Level 3 (L3) autonomy in 2025 and reach Level 4 (L4) in low-speed scenarios next year.
XPeng’s L3 + L4 Autonomous Driving
During the latest earnings call in March, XPeng’s chief executive stated to “believe that in the second half of this year, XPeng will be the first in China to offer a smart driving experience that boasts software capabilities and user experience equivalent to L3 autonomous driving.”
“L3 autonomy will generate strong user demand and exceptional user loyalty, marking an iPhone 4 moment for AI-defined cars,” Xiaopeng added.
On the call, the CEO also highlighted that “from a technology standpoint”, Tesla and XPeng are “the only two companies worldwide that are capable of providing globally reliable AI smart driving experience […] using a single software suite across the vehicle models.”
The company aims to compete “alongside the world’s leading AI autonomous driving companies in both the Chinese and global markets.”
AI Development
According to the Guangzhou-based company, the ‘Cloud Model Factory’ handles tasks like pre-training, post-training, model distillation, and vehicle-side model training. This system runs on 10 EFLOPS of computing power, maintains over 90% cluster efficiency, and completes training cycles in about five days.
Developed last year, XPeng’s AI infrastructure includes the first large-scale intelligent computing cluster in China’s automotive industry. This foundation supports the new XPeng World Base Model, which will power the company’s broader AI ecosystem—including smart cars, AI robots, and flying vehicles.
XPeng has already collected 20 million video clips to train its models, with plans to reach 200 million by the end of the year.
