Author | Ben Yi, Editor | De Xin
At the Technology Day, Zhu Jiangming announced that on March 9, 2026, Leapmotor successfully ran a demo version of its world model-based autonomous driving system.
Seventeen days later, the A10 was launched. Leapmotor used the term "World Model ADAS" in its external materials, emphasizing that this architecture does not require extreme computing power.
Five months later, on August 19, a paper titled "DA-WAM: Decision-Aligned Future Latents for Driving World Models" was uploaded to arXiv, with Leapmotor listed as one of the affiliations. This is Leapmotor's first paper related to world models for autonomous driving.
On the evening of September 16, at Leapmotor's 2026 Technology Day, the LWM (Leapmotor World Model) was officially released.
In August, we analyzed seven autonomous driving-related papers from Huawei, Li Auto, XPeng, and Xiaomi. Our conclusion is that although each company has its own focus in choosing academic paths, none of them has truly denied the concept of "predicting the future." In other words, integrating predictive world models into autonomous driving has become an industry consensus.
Leapmotor is no exception.
Taking the opportunity of the 2026 Technology Day, we hope to explore from both academic and mass-production perspectives what technical choices Leapmotor's world model has made, what the costs of these choices are, and how far it has progressed toward its promise of a "world model autonomous driving system available for under CNY 100,000."
Internal "Tian Ji's Horse Race" Strategy: Version 4.0 Emerges Victorious
Leapmotor internally named this world model-driven autonomous driving system "AD 4.0".
In October 2025, Leapmotor simultaneously pursued three internal technical routes. Version 2.0 is the already mass-produced two-stage end-to-end system, referred to as CNAP 2.0 in the launch event PPT. Version 3.0 retains the two-stage architecture but increases the proportion of models and reduces reliance on rules. Version 4.0 is more aggressive: a one-stage end-to-end world model that directly takes images as input, supplemented by a minimal set of rules to output decisions.
Versions 3.0 and 4.0 were each advanced by a team of over 20 members. Zhou Hongtao, Senior Vice President of Leapmotor, described it as a rare period of confusion for the company, where it was difficult to choose between the two directions through repeated discussions.
The first line of code for 4.0 was written in late 2025. In March this year, the two algorithms underwent comparative testing. Leapmotor invited third-party institutions and user groups to score them across subjective and objective dimensions. The result showed that while the in-car experience of 3.0 was superior to 2.0, it still could not beat 4.0.
The key metric provided by Zhou Hongtao is the comprehensive takeover rate, with 4.0 achieving only one-third of that of leading automakers. This metric includes two scenarios: efficiency takeovers and safety takeovers. The former includes situations where the autonomous driving system cannot bypass a leading vehicle, while the latter involves risks such as collisions or running red lights.
After this test, 3.0 was halted, and internal resources were concentrated on 4.0. Recalling this experience, Zhou Hongtao said that decisions that were not good enough, just like AD 3.0, were also terminated in a timely manner. His criticism of the two-stage end-to-end approach was straightforward: "It's like the telephone game played during team building; the people at the end receive much less information, requiring multiple confirmations, a process that increases latency."
Supporting this internal "Tian Ji's Horse Race" strategy was a major organizational restructuring.
In 2024, Leapmotor integrated relevant departments into the Intelligent Technology Research Institute, coordinated by Zhou Hongtao. The staff size approached 700, computing power resources doubled, and the number of training servers reached nearly 600. Currently, the scale of Leapmotor's autonomous driving team has exceeded 800 members.
In the same year, Leapmotor's Product and Technology Strategy Committee set a goal to join the first tier of autonomous driving by 2026, and to advance to L3 in 2027 and 2028. Since September 2025, Zhu Jiangming has designated autonomous driving as his number one project, holding meetings with the autonomous driving management team every half month.
The training data for the 4.0 system algorithm comes from Leapmotor's previous vehicle models. LiDAR takes on another identity during the training phase. LiDAR data from B and C series models are used as ground truth to supervise the model's pre-training. The D series is larger, with different sensor installation positions, providing diverse samples. After cleaning and filtering, approximately 30 million video clips entered the pre-training phase of 4.0.
In the edge-cloud architecture disclosed by Leapmotor, the cloud-based foundation model handles understanding, prediction, and scenario generation, while the vehicle-side model is deployed on various computing power platforms such as 200, 640, and 1280 TOPS | Image Source: Leapmotor 2026 Technology Day Launch Event
Abandoning BEV: Leapmotor Provides a Third Answer
In the presentation introducing the world model, there are only two quantifiable claims: "achieving the same model performance with only one-fifth of the computing power" and "reducing the amount of running code by 90%."
The first claim does not specify the benchmark it is compared against, nor does it explain what metrics are used to measure "the same performance." Throughout the presentation, there were no public benchmarks, no latency figures, no ablation comparisons, and no mention of any papers.
However, one slide in the launch event PPT stating "abandoning the traditional BEV perspective" successfully caught our attention.
Leapmotor lists "abandoning the traditional BEV perspective" and "reducing the amount of running code by 90%" as the core designs of the world model, explaining them as the first principles of mimicking human driving. | Image Source: Leapmotor 2026 Technology Day Launch Event
Regarding the perspective on BEV, different autonomous driving teams have varying voices. We actually left a question in the article we published in August. At that time, XPeng's X-Mind technical report mentioned using BEV as the base for an abstract world, compressing the future 12 frames into 96 tokens to represent the BEV layout. Li Auto's SparseWorld-TC explicitly bypasses the BEV intermediate representation, arguing that the explicit geometric constraints of BEV would limit the flexible interaction of spatiotemporal features. The judgments of the two companies are opposite, so we said we would wait and see more mass-production results.
Leapmotor happens to provide a third answer, but for reasons completely different from the other two. According to Feng Mingyue, head of Leapmotor's world model autonomous driving, Leapmotor's abandonment of BEV is mainly based on considerations across three dimensions.
First, BEV solves the dimensionality problem. In an image, a one-pixel difference between the top and bottom pixels might actually represent a ten-meter difference in reality due to perspective. BEV flattens this out.
Second, humans do not have a god's-eye view when driving; the perspective effect itself is useful. He gave an example of a cut-in maneuver. Traditional two-stage systems require a dedicated classification head for cut-ins to determine the heading angle and state of the oncoming vehicle. However, at high speeds, when a vehicle cuts into the ego lane, its changes in heading angle and position are very subtle, providing a weak gradient signal to the model. In contrast, in an image with perspective, when a vehicle covers your front from the right side, the pixel changes are extremely large.
Finally, the BEV transformation itself is quite computationally expensive. For Leapmotor, which pursues cost and efficiency, this is an intolerable technical "dead weight," so discarding it is perfectly reasonable.
Therefore, the March demo achieved "perception without relying on BEV." "When humans drive, they do not construct a BEV bird's-eye view in their minds, so the 4.0 solution is determined to use pure image input to directly output trajectories, without redundant transformations," Zhou Hongtao pointed out.
However, there is another phrase on the launch event PPT that does not quite stand up to scrutiny. "Eyes directly connected to hands and feet," with the small text below reading "directly outputting steering wheel and throttle signals." This is inconsistent with what Zhou Hongtao said in the interview.
According to Zhou Hongtao, the model itself does not directly output control signals; it only generates the expected driving trajectory for the next 4 seconds. On this basis, the R&D team integrates a minimal rule layer to translate trajectory intentions into throttle and steering wheel commands. The core information document provided officially by Leapmotor explains it more specifically, stating that "rules are only used as a fallback for traffic regulations such as traffic lights, reducing the code volume by 90%."
We can understand that the rule layer is indeed designed to be very thin. Zhou Hongtao also emphasized that post-processing must align with the model's intentions and cannot rely on the rule experience from the two-stage era. For safety considerations, "no matter how many 9s the model's accuracy reaches, the 1 representing unknown scenarios always exists, and ultimately, a small amount of rules must still serve as a fallback."
The launch event summarized LWM as "one brain, seamlessly connecting perception to action," and used the communicative expression of "eyes directly connected to hands and feet"; however, according to Zhou Hongtao, the model outputs future trajectories and still requires a small amount of rules for translation and fallback | Image Source: Leapmotor 2026 Technology Day Launch Event
But "directly outputting steering wheel and throttle signals" is factually incorrect. We guess this was likely done to emphasize "one brain, seamlessly connecting perception to action," with the marketing communications team removing the descriptions regarding "rules."
Hidden Details in the World Model Paper That Leapmotor Didn't Mention
The DA-WAM paper was completed in collaboration between The Hong Kong University of Science and Technology (Guangzhou) and Leapmotor. The designated project leader, Lang Zhang, is affiliated with Leapmotor.
Interestingly, we also found another world model paper authored by Lang Zhang (LiSTAR: Ray-Centric World Models for 4D LiDAR Sequences in Autonomous Driving), uploaded to arXiv in November 2025. In that paper, his affiliation is Li Auto, and co-authors include members of Li Auto's autonomous driving team.
The DA-WAM code repository is hosted under a GitHub organization called LeapWM. One is for robotics, LeapBot-WA, another is called levelk-wam, and the other two are currently empty repositories. Level-K refers to hierarchical reasoning in game theory.
The problem DA-WAM aims to solve is actually the same thing discussed at Leapmotor's Technology Day launch event this time.
It posits that the upper limit of a world model's value depends on how directly the prediction affects the scoring of candidate trajectories. Therefore, it predicts a separate future for each candidate trajectory and then uses this future to score that specific trajectory.
DA-WAM argues that multiple candidate trajectories sharing the same future will cause "prediction-action mismatch"; it predicts future representations separately for each candidate trajectory and then applies the corresponding future to score that trajectory | Image Source: DA-WAM, Figure 1
The paper includes a matching ablation table where the training data, initialization, candidate generator, training schedule, and evaluation protocol are all fixed, with only the connection method for future prediction changed.
A standard end-to-end planner that does no future prediction at all scores 93.31. Having all candidate trajectories share the same predicted future yields 92.81. Connecting only the hidden state of the current moment yields 93.25. Assigning a separate future to each candidate trajectory yields 93.46. Adding safety-related hard negative sample supervision yields 93.68.
After controlling for conditions such as training data, initialization, candidate generator, and evaluation protocol, DA-WAM improves from 93.31 without future prediction to 93.68; when sharing the same future, the score instead drops to 92.81. The table above also lists its 87.7 EPDMS on NAVSIM-v2 | Image Source: DA-WAM, Table 2, Table 3
The maximum difference among the five numbers is less than one point. Completing the full set improves the score from 93.31 to 93.68, an increase of 0.37 points.
Among these five numbers, the second one is the most noteworthy.
When integrating the world model, if all candidates share one future, the score is half a point lower than doing no future prediction at all. The paper's own explanation is that sharing the future causes a mismatch between prediction and action; the model learns an averaged-out representation, which instead weakens the discrimination among candidates.
This sentence can actually directly answer the question, "Are world models actually useful?" They are useful, but connecting them incorrectly can lead to negative returns.
There is also a key piece of data in the sub-items of this matching ablation table. From no future prediction to the full version, safety-related metrics have increased across the board: the collision-free rate rises from 98.45 to 99.11, drivable area compliance from 98.27 to 98.88, and time-to-collision from 95.48 to 96.81. In the same group, the ego vehicle's progress drops from 91.36 to 89.97.
Therefore, what the world model buys is safety, and what it pays for is traffic efficiency. The car drives more steadily, but also more conservatively.
DA-WAM demonstrates trajectory selection in three types of scenarios: large-angle left turns, congestion, and yield conflicts. The paper states that its method avoids the collision and TTC risks present in the baseline approach in the latter two scenarios | Image Source: DA-WAM, Figure 4
"Between safety and efficiency, we will definitely prioritize safety first. If a god's-eye view tells you that this driver is absolutely safe, you would certainly dare to ride with them. So if I just drive a bit slower, what's the harm in being a little slower?" said Feng Mingyue.
On NAVSIM-v1, DA-WAM's 93.7 leads the second place in the paper's comparison table by 0.2 points. On NAVSIM-v2, it achieves 87.7 EPDMS.
DA-WAM achieves 93.7 PDMS in the NAVSIM-v1 comparison listed in the paper, 0.2 points higher than the second place in the table. The study is configured with two frames of historical input from the front-facing camera, 32 candidate trajectories, and 0.5 seconds of future prediction | Image Source: DA-WAM, Table 1
Looking beyond the paper, the gap is even smaller. Drive-JEPA, involving XPeng, self-reported v1 at 93.3 and v2 at 87.8 in January this year. DreamerAD, a collaboration between the Institute of Automation, Chinese Academy of Sciences, and Chongqing Changan Technology, self-reported v2 at 87.7 in March. The three are squeezed between 87.7 and 87.8 on v2.
What must be clarified here is that DA-WAM is not the foundation model for Leapmotor's mass-produced 4.0 autonomous driving system.
The V-JEPA 2.1 initialized visual encoder used in the paper involves Leapmotor and The Hong Kong University of Science and Technology using Meta's V-JEPA 2.1 to initialize the visual encoder, followed by low-rank fine-tuning. It was trained for 20 epochs on 8 GPUs, with input consisting of only two frames of history from the front-facing camera, predicting 0.5 seconds into the future, and generating 32 candidate trajectories.
According to Zhou Hongtao, the mass-produced world model autonomous driving system is built on a cloud-based foundation model with approximately 128 billion parameters. The training data comprises about 30 million video clips, and the vehicle-side outputs a 4-second trajectory. Leapmotor states that the response time of the vehicle-side model on a 200 TOPS platform is approximately 200 milliseconds.
In comparison, the two differ by several orders of magnitude. This DA-WAM paper serves as a public slice demonstration of Leapmotor's self-developed world model route, rather than being 100% the theoretical foundation for this mass-produced 4.0 system.
Algorithms Trained by a Security Surveillance Company
Why did Leapmotor choose a path that saves computing power? We can find the answer in the company's personnel structure.
Zhu Jiangming is one of the founders of Dahua Technology and founded Leapmotor in 2015. Zhou Hongtao joined Dahua after graduating in 2001, followed Zhu Jiangming to join Leapmotor in 2016, and is now the Senior Vice President of Leapmotor, coordinating the Intelligent Technology Research Institute.
Wang Yaonong, the most frequently appearing name in Leapmotor's publicly published series of autonomous driving papers, and another frequent author, Hu Laifeng (Laifeng Hu), were still at Zhejiang Dahua Technology in 2020. That year, they submitted a technical report for the EPIC-Kitchens Object Detection Challenge using email addresses with the @dahuatech.com suffix. In March 2023, the affiliation for papers by these same two individuals changed to Zhejiang Leapmotor Technology.
Public reports show that Wang Yaonong serves as the Vice President of Leapmotor Technology and head of autonomous driving. Laifeng Hu's real name and position are unknown.
From the CEO to the head of autonomous driving and then to the paper authors, the entire lineage is rooted in Dahua's security surveillance visual algorithm team. "Nothing new is created out of thin air; it is all about summarizing the shortcomings of the previous solutions," said Feng Mingyue.
This explains why the theme of Leapmotor's autonomous driving papers over the past three years has consistently been "efficiency." The 2023 GAM focused on efficiency optimization for point cloud analysis, while the 2024 ADMap and 2025 FastMap addressed jitter and decoder redundancy in online vectorized maps. These individuals did not enter from the autonomous driving academic circle; the security camera business inherently involves doing the most visual computing on the cheapest hardware.
In June this year, at the launch communication meeting for the new C series, Zhu Jiangming said, "AI autonomous driving chips are somewhat oversupplied." His reasoning is that there are over a dozen AI autonomous driving chips on the market, yet the annual demand is only 10 to 20 million units. As for self-developed chips, he said they would consider making their own only when Leapmotor becomes Toyota.
At this year's Technology Day launch event, he stated this judgment more directly: "Algorithm innovation is more important than piling on computing power."
At the launch event, Leapmotor reviewed its electronic product roadmap from automotive-grade autonomous driving chips to the 8797 cockpit-driving integrated central domain controller | Image Source: Leapmotor 2026 Technology Day Launch Event
Where did the saved money go? Zhou Hongtao provided a figure. The cost of Leapmotor's EEA 4.0 architecture is only 40% of that during the 1.0 era. The reduction in underlying costs freed up budget for BOM expenditures on LiDAR, cameras, and autonomous driving chips. This is the commercial premise for bringing the world model to the A-platform vehicle models priced around CNY 100,000.
However, according to official data, as of the end of August, 350,000 vehicle models equipped with LiDAR and high-level autonomous driving chips have been delivered. Leapmotor has promised all LiDAR version users a free upgrade to the latest version with lifetime free usage. For the earliest batch of over 10,000 LiDAR version users, due to inconsistent hardware platforms, algorithm migration and reconstruction will cost at least CNY 50 million.
Over 10,000 vehicles, CNY 50 million, averaging several thousand yuan per vehicle. This is the real money Leapmotor has spent for the concept of "technology equity."
Media Drives the 640, Leapmotor Sells the 200
A week before the launch event, Leapmotor invited some media to experience this world model autonomous driving system in advance in Hangzhou. In the currently public real-world test videos, the test drive vehicles are all D19.
One media representative said, "Today we only drove with a computing power of 640 TOPS and already achieved this, so I am quite looking forward to what it will be like later if both chips are activated." But in fact, the focus of Leapmotor's Technology Day this year is "world model high-level ADAS available for under CNY 100,000," which corresponds to the 8650 chip on the A10 with 200 TOPS of computing power.
At the launch event, Leapmotor made "world model high-level ADAS available for under CNY 100,000" its core promise, with the actual performance on the corresponding low-computing platform yet to be verified by mass-produced vehicles | Image Source: Leapmotor 2026 Technology Day Launch Event
Several media members mentioned impressive scenarios during the test drive. Some said this D19 could snatch half a second when the light turns from red to green, "This is the first car I've used with ADAS that can snatch half a second." It doesn't brake early when the countdown reaches 5, 4, 3, but carries the green light through just like a human.
One media member captured a scenario of passing vehicles in a narrow alley, with three large vehicles and a Mazda with its rear end sticking out coming from the opposite direction. The car did not recklessly squeeze in but kept adjusting its steering until there was space. A passenger in the car said it could deduce that if it plunged in, no one would be able to move.
Another media member encountered an intersection with a right-turn arrow light accompanied by the text "Right-turning vehicles prohibited." The system waited until the text turned off before proceeding. Leapmotor's explanation is that the model does not understand Chinese characters; it relies on the camera to scan the patterns of the arrow and text to make judgments, with logic similar to Tesla recognizing stop signs.
It is said that after exiting navigation, there is also a roaming mode where the model decides on its own whether to go straight, turn left, or turn right, rather than a hard-coded right-turn loop based on a few rules.
Opinions on Leapmotor's world model-driven high-level ADAS also vary. Some said, "There is still very obvious room for growth, such as hesitation, stuttering, and handling in extremely complex scenarios." "I wouldn't say it's already the strongest in the industry," because this is still the first version in the internal testing phase. Others expressed that "current ADAS is on a level playing field for everyone, and Leapmotor can be added to the list."
Leapmotor lists typical capabilities of LWM including defensive driving, low-speed creeping, mapless roaming, three-point turns, and voice-activated roadside parking | Image Source: Leapmotor 2026 Technology Day Launch Event
A very well-known industry expert gave this system a T0-level evaluation, adding that it is a surprise, but not a super massive one, because he has tried many high-level ADAS models and everyone has done quite well. He then asked a question that many people who watched Leapmotor's Technology Day release this year might want to ask.
"I am very curious, after distilling a version that can run on a relatively low-computing processor, how much of today's level can be retained? The manufacturer says it can retain 100%, but I still have my doubts."
Currently, the vehicles driven by the media are basically D19. This car is equipped with two Snapdragon 8797 chips, one for the cockpit and one for ADAS, with a single chip's sparse computing power at 640 TOPS. But Leapmotor promises that the same system will be deployed on the CNY 100,000-class models with 200 TOPS.
"Currently, our world model is developed on the Qualcomm 8797 chip, but next year it will definitely be deployed on the 8650 chip, and even on a 100 TOPS computing platform, we must achieve a relatively good performance. This is also the goal we have set for the team," Zhou Hongtao stated in a media interview.
Image Source: Leapmotor 2026 Technology Day Launch Event
According to the upgrade sequence announced on the launch event PPT, the 350,000 existing LiDAR version users will complete the world model autonomous driving update in four batches, with the A10 and A05 scheduled for the fourth quarter of 2027. However, Feng Mingyue revealed, "Before the middle of 2027, we will OTA all existing 8650 models." The A10 uses exactly the 8650.
Regardless of which timeline is used, from this launch event to the moment that "under CNY 100,000" vehicle model actually receives the world model, there may still be a gap of nine to fifteen months.
Embodied AI is in the Works, But No Robots Were Walked on Stage
At the end of the Technology Day release, someone asked Zhu Jiangming why Leapmotor is not getting into embodied intelligence.
His answer was that Leapmotor is in the best position to do this, but the company's philosophy is not to hype concepts. Currently, they have only made small-scale investments in embodied intelligence and are in a following state. He admitted that Leapmotor has already developed a fully self-developed robot. Colleagues suggested bringing the robot on stage for a walk, but he rejected the idea.
"A robot that can just be walked around on stage doesn't count as a real achievement; we only show off robots that can make money."
In a media interview, Zhou Hongtao provided a more detailed introduction to Leapmotor's investments in embodied intelligence. Currently, the Leapmotor Intelligent Technology Research Institute has five departments: data closed-loop, end-to-end models, mass production development, computing power platforms, and the fifth is embodied intelligence. A robot prototype has been developed, with technical reserves in both humanoid and wheeled directions. The short-term goal targets two scenarios: factory automation and logistics transportation, aiming to reverse-strengthen the overall vehicle's cost competitiveness by improving production line and warehousing efficiency.
In fact, Leapmotor's public record of entering the embodied intelligence track is even earlier than what Zhou Hongtao told the media.
On July 27, a paper called LeapBot-WA was uploaded to arXiv, focusing on robotic arm manipulation. Experiments were conducted on LIBERO and RoboTwin 2.0, and real-machine migration was completed. The authors include Wang Yaonong, Jiachao Liu, and Feng Mingyue, the same group of people who worked on DA-WAM, with the code hosted under the same LeapWM organization. At the end of July, Huzhou Lingsheng Precision Manufacturing Co., Ltd., wholly established by Zhejiang Lingsheng Power under Leapmotor, was founded with a registered capital of CNY 210 million. Its business scope includes industrial robot manufacturing and intelligent robot R&D.
What Zhu Jiangming said on stage was small-scale investment and following, yet in the exclusive interview, Zhou Hongtao clearly detailed the staffing and prototype progress of the embodied intelligence department. Coupled with the already published papers and the registered company, it truly feels like they are trying to hide something but inadvertently drawing more attention to it.
However, this is roughly the same meaning as what Zhu Jiangming said when discussing factory automation at the end of 2025. Investments in automation equipment must be able to recover the investment through saved labor costs within three years; if not, they prefer to hold off. Perhaps Leapmotor's robots are still stuck on scenario realization, and the commercialization math doesn't quite add up yet.
References
Leapmotor 2026 Technology Day on-site speech and official core information documents, September 16, 2026
Ni Ping, "Exclusive Interview with Leapmotor's Zhou Hongtao: Sprinting to the Top Tier of Autonomous Driving in 2026, Leapmotor's Progress Bar is Faster Than Expected," Leiphone, September 17, 2026
DA-WAM: Decision-Aligned Future Latents for Driving World Models, arXiv:2608.19085
LeapBot-WA: World-Anchor Action Models via Predictive Latent Alignments, arXiv:2607.23969
GAM: Gradient Attention Module of Optimization for Point Clouds Analysis, AAAI 2023, arXiv:2303.10543
ADMap: Anti-disturbance Framework for Vectorized HD Map Construction, ECCV 2024, arXiv:2401.13172
FastMap: Fast Queries Initialization Based Vectorized HD Map Reconstruction Framework, arXiv:2503.05492
DHARI Report to EPIC-Kitchens 2020 Object Detection Challenge, arXiv:2006.15553
LiSTAR: Ray-Centric World Models for 4D LiDAR Sequences in Autonomous Driving, arXiv:2511.16049
Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving, arXiv:2601.22032
DreamerAD: Efficient Reinforcement Learning via Latent World Model for Autonomous Driving, arXiv:2603.24587
V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning, arXiv:2603.14482
Stellantis Media, Leapmotor A10 Launch in China, March 26, 2026
Leapmotor Automobile 2026 Interim Report, Stock Exchange of Hong Kong
Media public real-world test videos, September 2026
Feng Mingyue, head of Leapmotor's world model autonomous driving, media interview video, September 2026