- MindVLA will redefine autonomous driving in the same way that the iPhone 4 redefined smartphones, said Li Auto CEO.
- Li Auto management said earlier this month that it initiated R&D on the next-generation VLA smart driving large model, which will be released alongside the Li i8.
Li Auto (NASDAQ: LI) has unveiled its next-generation autonomous driving architecture, MindVLA (Visual-Language-Action), which aims to move toward truly autonomous driving.
The Chinese carmaker’s head of autonomous driving technology development, Jia Peng, unveiled the autonomous driving architecture at the ongoing Nvidia GTC 2025 event, according to an announcement by Li Auto.
“MindVLA is a Visual-Language-Action large model, but we prefer to call it a ‘robot large model,'” Li Xiang, founder, chairman, and CEO of Li Auto, said in a Weibo post.
The model unifies spatial, linguistic and behavioral intelligence in a single model, giving autonomous driving systems the ability to sense, think and adapt to their environment, and is the most important step on Li Auto’s road to L4 autonomous driving, Li said.
MindVLA’s ability to empower autonomous driving with human-like driving capabilities will redefine autonomous driving in the same way the iPhone 4 redefined smartphones, Li said.
“MindVLA is here, is truly autonomous driving still far away?” Li wrote on Weibo.
During Li Auto’s fourth-quarter 2024 earnings analyst call on March 14, the company’s management said it initiated research and development on the next-generation VLA smart driving large model, which will be released alongside the Li i8.
The Li i8 is Li Auto’s first all-electric SUV (sport utility vehicle) model and is expected to be launched in July.
MindVLA will transform the car from a mere means of transportation to a full-time driver, Jia said at the Nvidia GTC 2025 event, adding that Li Auto hopes MindVLA will give the car human-like cognitive and adaptive capabilities, transforming it into a smart agent capable of thinking.
Li Auto designed and trained a large language model base model suitable for MindVLA from scratch, using the MoE hybrid expert architecture and introducing the Sparse Attention mechanism, according to Jia.
This design ensures that the model size grows without decreasing the reasoning efficiency on the user side, according to Jia.
MindVLA uses diffusion to decode action tokens into optimized trajectories and generates joint modeling that incorporates trajectory predictions of other vehicles through the vehicle’s own behavior, improving gaming capabilities in complex traffic environments, according to the company.
Based on Li Auto’s in-house developed world model, the autonomous driving architecture can build simulated environments that are close to the real world, according to Jia.
Li Auto to launch Li i8 electric SUV in Jul, shares 1st teaser video
