Leapmotor Claims First-Tier ADAS With Self-Developed World Model

At its 2026 Technology Day, Leapmotor announced the LWM (Leapmotor World Model) intelligent driving system, marking what the company called a jump into the "first tier" of advanced driver-assistance systems. The automaker played a continuous dashcam-style test video featuring urban driving, narrow-road passing, unprotected intersections, and navigation-free roaming — scenarios Leapmotor said demonstrated its system's defensive-driving and prediction capabilities.

Leapmotor also committed to rolling out the world-model system beginning in Q1 2027, starting with its LiDAR-equipped models and extending to vehicles priced under 100,000 RMB (approx. $13,889), a price band where such technology has been scarce.

A Year of Heavy Investment

CEO Zhu Jiangming said intelligent driving became his top priority last September, with bi-weekly management reviews. By March 9, 2026, a demo version of the world-model system was running on vehicle hardware, which Leapmotor said placed its algorithms in the first tier.

The smart-driving team now numbers over 800 people. According to electronics-line head Zhou Hongtao, Leapmotor achieved full in-house stack development and can iterate up to eight algorithm versions per day.

Research Paper Reveals the Approach

Leapmotor pointed to a paper co-authored with Hong Kong University of Science and Technology, published in August 2025, titled DA-WAM: Decision-Aligned Future Latents for Driving World Models.

The paper argues that a world model's planning value depends on how directly its predictions influence candidate-trajectory scoring. DA-WAM addresses this by coupling future prediction with decision-making: each candidate trajectory receives its own independently predicted future latent state, which a factorized scorer then evaluates alongside the trajectory.

Key technical moves include an online V-JEPA 2.1 encoder paired with an exponential moving average target encoder, ensuring predictive representations adapt throughout planner optimization rather than freezing after pre-training. The system also combines matched future goals with safety-critical hard negative samples during training.

In benchmark results reported in the paper, DA-WAM scored 93.7 PDMS on NAVSIM-v1 (camera-only city driving with navigation), trailing human drivers at 94.8 PDMS. On NAVSIM-v2 it achieved 87.7 EPDMS, Leapmotor said it outperformed other learned planners in both tests.

Rollout Schedule and Hardware Diversity

The Q1 2027 upgrade will cover 350,000 Leapmotor vehicles equipped with LiDAR. The initial wave includes the Leapmotor D19 and the 2027 C10 and B10. The 2027 C16 and B01 are slated for Q2 2027, with remaining models moving through Q3 and Q4.

A complication is hardware diversity across the fleet. The D19's higher-trim versions use dual Qualcomm SA8797P chips delivering 1,280 TOPS, while LiDAR-equipped 2027 B10 models use the Qualcomm SA8650 at 200 TOPS. Zhu acknowledged that rebuilding the algorithm for inconsistent hardware platforms would cost at least 50 million RMB (approx. $6.94 million) and significant engineering resources.

The challenge mirrors a broader industry hurdle. In June 2026, NIO announced a world-model update covering over 700,000 vehicles, highlighting that it had co-developed its software for both generic and in-house chips to keep experiences consistent across platforms.

What Comes Next

Leapmotor's technology day also showcased a new vehicle architecture, a central domain-control system, and updated battery and electric-drive components. The company's long-standing reputation as an ADAS follower is now being challenged by its push toward self-developed world-model driving.

Whether Leapmotor can deliver a uniform experience across divergent hardware remains an open question — one the company will have to answer as the Q1 2027 rollout begins.