Author | Ben Yi

Editor | Dexin

Zhu Jiangming said at the Technology Day that on March 9, 2026, Leapmotor successfully developed a demo version of its World Model Intelligent Driving System.

Seventeen days later, the A10 was launched, and 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 among the affiliated institutions. This was Leapmotor's first paper related to intelligent driving world models.

On the evening of September 16, during Leapmotor's 2026 Technology Day, the LWM Leapmotor World Model was officially released.

Image Source | Leapmotor 2026 Technology Day Launch Event

In August, we dissected seven papers related to intelligent driving from Huawei, Li Auto, XPENG, and Xiaomi. Our conclusion was that while each company had its own focus in academic path selection, none truly denied the importance of 'predicting the future.' In other words, integrating predictive world models into autonomous driving has become an industry consensus.

Leapmotor is no exception.

Taking advantage of the 2026 Technology Day, we hope to examine, from both academic and mass-production perspectives, what technical choices Leapmotor has made with its world model, what the costs of these choices are, and how far along it is in delivering on its promise of a world model intelligent driving system 'available for under 100,000 yuan.'

I. Internal 'Horse Racing,' with 4.0 Emerging Victorious

Internally, Leapmotor named this world model-driven intelligent driving system 'Intelligent Driving 4.0.'

In October 2025, Leapmotor simultaneously pursued three internal routes. Route 2.0 was a two-stage end-to-end system already in mass production, referred to as CNAP 2.0 in the presentation slides. Route 3.0 continued with a two-stage architecture but increased the proportion of models and reduced reliance on rules. Route 4.0 was more aggressive, featuring a one-stage end-to-end world model that directly input images and supplemented decision-making with a small number of rules.

Image Source | Leapmotor 2026 Technology Day Launch Event

Routes 3.0 and 4.0 were each advanced by teams of over 20 people. Zhou Hongtao, Leapmotor's Senior Vice President, described this period as one of rare uncertainty for Leapmotor, with both directions being difficult to choose between in repeated discussions.

The first line of code for 4.0 was written at the end of 2025. In March of this year, the two algorithms underwent PK testing, with Leapmotor inviting third-party organizations and groups to score them based on both subjective and objective dimensions. The results showed that while 3.0 offered a better experience than 2.0 after implementation, it still could not match 4.0.

Zhou Hongtao highlighted the key metric of comprehensive takeover rate, with 4.0 having only one-third that of leading automakers. This metric includes both efficiency takeovers (e.g., when intelligent driving cannot bypass a vehicle ahead) and safety takeovers (e.g., when there is a risk of collision or running a red light).

After this test, Route 3.0 was discontinued, and internal resources were concentrated on 4.0. Reflecting on this experience, Zhou Hongtao said that decisions that did not perform well, like Intelligent Driving 3.0, were promptly discontinued. He was straightforward in his criticism of the two-stage end-to-end approach: 'It's like the telephone game played during team-building activities, where the information received by the later participants is significantly diminished, requiring multiple confirmations and increasing delays.'

This internal 'horse racing' approach was supported by a major organizational restructuring.

In 2024, Leapmotor consolidated its relevant departments into the Intelligent Technology Research Institute, led by Zhou Hongtao, with a team of nearly 700 people. Computing power resources tripled, and the number of training servers approached 600. Currently, Leapmotor's intelligent driving team has grown to over 800 people.

Image Source: Leapmotor 2026 Technology Day Launch Event

In the same year, Leapmotor's Product Technology Strategy Committee set a goal to join the first tier of intelligent driving by 2026 and advance to L3 by 2027 and 2028. Since September 2025, Zhu Jiangming has made intelligent driving his top project, holding meetings with the intelligent driving management team every two weeks.

The training data for this 4.0 system algorithm came from Leapmotor's past vehicle models. LiDAR played another role during training. LiDAR data from B and C models was used as ground truth to supervise the model's pre-training. The D series, being larger and having sensors installed in different positions, provided diverse samples. After cleaning and screening, approximately 30 million video clips entered the pre-training stage for 4.0.

II. Abandoning BEV, Leapmotor Offers a Third Approach

In the presentation on the world model, there were only two quantifiable claims: 'achieving the same model effect with only one-fifth of the computing power' and 'reducing the amount of running code by 90%.'

The former did not specify who the comparison was made against or what metrics were used to measure 'the same effect.' There were no publicly disclosed benchmarks, latency figures, ablation comparisons, or references to any papers throughout the presentation.

However, one slide in the presentation PPT caught our attention: 'Abandoning the traditional BEV perspective.'

Intelligent driving teams have different views on BEV. In our August article, we left a question unanswered. At that time, XPENG's X-Mind technical report mentioned using BEV as the base for abstracting the world, compressing the next 12 frames into 96 tokens to represent the BEV layout. Li Auto's SparseWorld-TC explicitly bypassed BEV intermediate representations, arguing that BEV's explicit geometric constraints limit the flexible interaction of spatiotemporal features. The judgments of the two companies were opposite, and we said we would wait for more mass-production results to see.

Leapmotor has now provided a third answer, but for reasons different from the first two companies. According to Feng Mingyue, head of Leapmotor's World Model Intelligent Driving, Leapmotor abandoned BEV primarily based on considerations from three dimensions.

First, BEV solves the dimensionality problem. On an image, the difference between pixels at the top and bottom can be ten meters in reality due to perspective. BEV flattens this out.

Second, humans do not have a god's-eye view when driving, and perspective itself is useful. He gave an example of a cut-in. In a traditional two-stage system, a separate classification head is needed to judge the heading angle and state of an oncoming vehicle. However, when a vehicle cuts into the lane at high speed, its heading angle and position changes are very slight, providing weak gradient signals to the model. In a perspective image, a vehicle moving from the right side to in front of you creates significant pixel changes.

Finally, BEV conversion itself is quite computing power-intensive. For Leapmotor, which pursues cost and efficiency, this is an intolerable technological 'chicken rib,' and abandoning it is only normal.

Thus, the March demo version 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 be pure image input with direct trajectory output, without redundant conversions,' Zhou Hongtao pointed out.

However, there is a statement on the presentation PPT that does not hold up well under scrutiny. 'Eyes directly connected to hands and feet,' with the small print stating 'direct output of steering wheel and throttle signals.' This contradicts what Zhou Hongtao said in the interview.

According to Zhou Hongtao, the model itself does not directly output control signals but generates the expected driving trajectory for the next 4 seconds. The research and development team then incorporates a minimal number of rule layers to translate trajectory intentions into throttle and steering wheel commands. The core information document provided by Leapmotor explains it more specifically, stating that 'rules are only used to comply with traffic regulations such as traffic lights, reducing code volume by 90%.'

We can understand that the rule layer is indeed designed to be thin, and Zhou Hongtao also emphasized that post-processing must align with the model's intentions and cannot rely on rule experiences from the two-stage era. For safety reasons, 'no matter how high the model's accuracy reaches, the 1 representing unknown scenarios always exists, and a small number of rules must ultimately serve as a backup.'

However, the statement 'direct output of steering wheel and throttle signals' is not factually accurate. We speculate that this was done to emphasize 'one brain doing it all in one go,' and the description about 'rules' was removed for market communication purposes.

III. The Paper on World Models Hides Details Leapmotor Didn't Mention

The DA-WAM paper was a collaboration between the Hong Kong University of Science and Technology (Guangzhou) and Leapmotor, with Lang Zhang, the project leader listed, affiliated with Leapmotor.

Image Source: First page of the DA-WAM paper

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 was Li Auto, and he co-authored it with members of Li Auto's intelligent driving team.

The code repository for DA-WAM is hosted under a GitHub organization called LeapWM, with one repository for the robot LeapBot-WA and another called levelk-wam, while the other two are currently empty. Level-K refers to hierarchical reasoning in game theory.

The problem DA-WAM aims to solve is essentially the same as what Leapmotor discussed at its Technology Day launch event.

It argues that the upper limit of a world model's value depends on how directly predictions affect the scoring of candidate trajectories. Therefore, it predicts a separate future for each candidate trajectory and uses that future to score the trajectory.

The paper includes a matching ablation table, with all training data, initialization, candidate generators, training plans, and evaluation protocols fixed, and only the method of future prediction varied.

An ordinary end-to-end planner that does not make any future predictions scores 93.31. Having all candidate trajectories share the same predicted future scores 92.81. Only incorporating the hidden state at the current moment scores 93.25. Each candidate trajectory having its own future scores 93.46. Adding supervision with safety-related difficult negative samples scores 93.68.

The largest gap among these five numbers is less than one point. Implementing the full set of methods increases the score from 93.31 to 93.68, a gain of 0.37 points.

The second number among these five is the most noteworthy.

When the world model is incorporated and all candidates share the same future, the score is half a point lower than when no future predictions are made at all. The paper's own explanation is that sharing the future causes a mismatch between predictions and actions, resulting in the model learning an averaged representation that weakens the distinction between candidates.

This statement can directly answer the question of 'whether world models are useful.' They are useful, but the wrong implementation can result in negative returns.

In the sub-items of this matching ablation table, there is also key data. From not making future predictions to the full version, safety-related metrics all improve, with the collision-free rate increasing 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 vehicle's progress drops from 91.36 to 89.97.

Thus, what the world model buys is safety, at the cost of Traffic efficiency (traffic efficiency). The vehicle drives more steadily but also more conservatively.

'Between safety and efficiency, we will always prioritize safety. If a god's-eye view tells you that the driver is definitely safe, you would dare to drive. So what if I drive a bit slower?' Feng Mingyue said.

On NAVSIM-v1, DA-WAM's score of 93.7 leads the second-place score in the paper's comparison table by 0.2 points. On NAVSIM-v2, it achieves 87.7 EPDMS.

Looking beyond the paper, the gap is even smaller. Drive-JEPA, in which XPENG participated, reported scores of v1 93.3 and v2 87.8 in January this year. DreamerAD, a collaboration between the Institute of Automation, Chinese Academy of Sciences, and Chongqing Changan Technology, reported v2 87.7 in March. The three companies are clustered between 87.7 and 87.8 on v2.

It must be clarified that DA-WAM is not the base model for Leapmotor's mass-produced 4.0 intelligent driving system.

The paper uses V-JEPA 2.1 to initialize the visual encoder, which Leapmotor and the Hong Kong University of Science and Technology initialized and then performed low-rank fine-tuning, training on 8 GPUs for 20 epochs. The input consists of two historical frames from the front-view camera, predicting the next 0.5 seconds and generating 32 candidate trajectories.

According to Zhou Hongtao, the mass-produced world model intelligent driving system is built on a cloud-based base model with approximately 128 billion parameters, trained on about 30 million video clips, and outputs 4-second trajectories on the vehicle side. Leapmotor claims that the vehicle-side model has a response time of about 200 milliseconds on a 200 TOPS platform.

In comparison, there are several orders of magnitude difference. DA-WAM's paper can be seen as a publicly disclosed slice of Leapmotor's self-developed world model route, not the theoretical basis for the 100% mass-produced 4.0 system.

IV. An Algorithm Taught by a Security Company

The reason why Leapmotor chose a computing power-saving route can be found in the company's personnel structure.

Zhu Jiangming is one of the founders of Dahua Corporation and established Leapmotor in 2015. Zhou Hongtao joined Dahua after graduating in 2001 and followed Zhu Jiangming to Leapmotor in 2016. He is now Leapmotor's Senior Vice President, overseeing the Intelligent Technology Research Institute.

The names that appear most frequently in Leapmotor's series of published intelligent driving papers are Wang Yaonong and another common author, Hu Laifeng (phonetic translation, Laifeng Hu), who were still at Zhejiang Dahua in 2020. That year, they submitted a technical report for the EPIC-Kitchens object detection challenge using an email address with the suffix @dahuatech.com. In March 2023, papers by the same two individuals listed Zhejiang Leapmotor Technology as their affiliation.

Public reports indicate that Wang Yaonong serves as Vice President of Leapmotor Technology and head of intelligent driving. Laifeng Hu's real name and position are unknown.

From the CEO to the head of intelligent driving and then to the paper authors, the entire line consists of personnel from Dahua's security visual algorithm team. 'Any new thing does not emerge out of thin air but is based on summarizing the shortcomings of the previous solution,' Feng Mingyue said.

This explains why the theme of Leapmotor's intelligent driving papers over the past three years has consistently been 'efficiency.' GAM in 2023 optimized the efficiency of point cloud analysis, while ADMap in 2024 and FastMap in 2025 addressed jitter and decoder redundancy in online vector maps. These individuals did not come from the academic circle of autonomous driving; the business of security cameras inherently involves performing the most visual computations on the cheapest hardware.

Zhu Jiangming said at the launch event for the new C series in June this year, 'AI intelligent driving chips are somewhat excessive.' His reasoning was that there are more than a dozen AI intelligent driving chips on the market, with annual demand of only 10 to 20 million units. As for self-developed chips, he said he would consider whether to make them himself only if Leapmotor were to become like Toyota.

At this year's Technology Day press conference, he put it more bluntly: "Algorithm innovation is more important than stacking computing power."

Where did the saved money go? Zhou Hongtao provided a figure. The cost of Leapmotor's EEA4.0 architecture is only 40% of that in the 1.0 era. Only with reduced underlying costs can budget be allocated for BOM expenditures on LiDAR, cameras, and intelligent driving chips. This is the commercial prerequisite for the World Model to be applied to A-platform models priced at the RMB 100,000 level.

However, according to official data, as of the end of August, 350,000 vehicles equipped with LiDAR and high-level intelligent driving chips have been delivered. Leapmotor has committed to providing all LiDAR version users with free upgrades to the latest version and lifetime free use. For the first batch of over 10,000 LiDAR version users, due to inconsistent hardware platforms, algorithm transplantation and reconstruction will cost at least RMB 50 million.

Over 10,000 vehicles, RMB 50 million—that's several thousand yuan spent per vehicle. This is the tangible investment Leapmotor has made for its philosophy of "technological empowerment for all."

Image source: Leapmotor 2026 Technology Day press conference

V. Media Drove the 640, While Leapmotor Sells the 200

A week before the press conference, Leapmotor invited some media to experience this World Model intelligent driving system in advance in Hangzhou. In the publicly available real-world test videos, the test vehicles are all D19 models.

One media representative said, "Today, we've achieved this with just 640 TOPS of computing power, so I'm quite looking forward to what happens when both chips are fully utilized." But in fact, the focus of this year's Leapmotor Technology Day is "high-level assisted driving with the World Model available for under RMB 100,000," which corresponds to the 8650 chip in the A10 with 200 TOPS of computing power.

Several media representatives mentioned memorable scenarios during the test drive. Someone said this D19 model could Took half a second (seize half a second) when the traffic light turned from red to green, "This is the first vehicle among the assisted driving systems I've used that can seize half a second." When the countdown reached 5, 4, 3, it didn't brake prematurely but passed through the green light like a human driver.

One media outlet captured a scenario of meeting oncoming vehicles in a narrow alley. Facing three large vehicles and a Mazda sticking out its rear, the car didn't Insert recklessly (rashly cut in) but kept adjusting its direction until there was space. The person in the same car said it could deduce that if it forced its way in, no one would get through.

Another media outlet encountered an intersection with a right-turn arrow light paired with "No Right Turn for Vehicles" text. The system waited until the text disappeared before proceeding. Leapmotor explained that the model doesn't understand Chinese text but relies on camera scanning of the arrow and text patterns to make judgments, similar to Tesla's logic for identifying 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 following a few fixed rules for right turns.

Opinions on Leapmotor's World Model-driven high-level assisted driving system vary. Some say, "There's still significant room for improvement, such as hesitation, stuttering, and handling in extremely complex scenarios." "I wouldn't say it's already the strongest in the industry," as this is still the first version in the internal testing phase. Others stated, "Currently, everyone is on a similar level with driving assistance, so Leapmotor can be added to the list."

A highly renowned industry expert gave this system a T0-level evaluation, calling it a pleasant surprise but not an overwhelming one, as he had tested many high-level assisted driving models, and they all performed well. He then asked a question that many might have after watching this year's Leapmotor Technology Day press conference:

"I'm very curious about how much of today's performance can be retained when a distilled version runs on a relatively low-computing-power processor. The manufacturer claims 100% retention, but I'm skeptical."

Currently, the media test drives are mostly on the D19 model. This vehicle is equipped with two Qualcomm 8797 chips—one for the cabin and one for assisted driving, each with a sparse computing power of 640 TOPS. However, Leapmotor has promised that the same system will be deployed on the 200 TOPS model priced at the RMB 100,000 level.

"Our World Model is currently developed on the Qualcomm 8797 chip, but next year, it will definitely be deployed on the 8650 chip, and we aim to achieve relatively good performance even with 100 TOPS of computing power. This is the goal we've set for our team," Zhou Hongtao said in a media interview.

Image source: Leapmotor 2026 Technology Day press conference

According to the upgrade sequence announced in the press conference PPT, the 350,000 existing LiDAR version users will receive updates to the World Model intelligent driving system in four batches, with the A10 and A05 scheduled for the fourth quarter of 2027. However, Feng Mingyue revealed, "Before mid-2027, all existing 8650 models will receive OTA updates." The A10 uses the 8650 chip.

No matter which timeline is considered, from this press conference to when the "available for under RMB 100,000" model truly gets the World Model, there may still be a gap of nine to fifteen months.

VI. Embodied AI Is in Development, But No Robots Were Shown on Stage

At the end of the Technology Day press conference, someone asked Zhu Jiangming why Leapmotor wasn't pursuing embodied AI.

His response was that Leapmotor is well-positioned to do so, but its philosophy is to avoid hype. Currently, it has only made small-scale investments in embodied AI, adopting a follower stance. He admitted that Leapmotor has developed a fully self-researched robot, and colleagues suggested showing it on stage, but he rejected the idea.

"A robot that can just walk around isn't impressive; only a robot that can make money is worth showing off."

In a media interview, Zhou Hongtao provided more specific details about Leapmotor's investments in embodied AI. Currently, Leapmotor's Intelligent Technology Research Institute has five departments: data closed loop (closed-loop), end-to-end models, mass production development, computing power platforms, and the fifth being embodied AI. Prototypes of robots have been developed, with technical reserves in both humanoid and wheeled directions. The short-term goals are focused on two scenarios: factory automation and logistics transportation, aiming to enhance production line and warehousing efficiency to strengthen the cost competitiveness of its vehicles.

In fact, Leapmotor's public foray into the embodied AI sector was recorded even earlier than what Zhou Hongtao told the media.

On July 27, a paper titled LeapBot-WA was uploaded to arXiv, focusing on robotic arm manipulation. Experiments were conducted on LIBERO and RoboTwin 2.0, with successful real-world deployment. The authors included Wang Yaonong, Jiachao Liu, and Feng Mingyue—the same team behind DA-WAM, with code hosted under the same LeapWM organization. In late July, Huzhou Lingsheng Precision Manufacturing Co., Ltd., a wholly-owned subsidiary of Zhejiang Lingsheng Power under Leapmotor, was established with a registered capital of RMB 210 million, covering industrial robot manufacturing and intelligent robot R&D.

What Zhu Jiangming said on stage about small-scale investment and following, while Zhou Hongtao clearly outlined the embodied AI department's structure and prototype progress in interviews, combined with the publicly available paper and registered company, gives a somewhat "here's the silver, but there's no thief" impression.

However, this aligns with Zhu Jiangming's statement in late 2025 regarding factory automation. He mentioned that investments in automation equipment must be recouped through labor cost savings within three years; otherwise, they would rather delay. Perhaps Leapmotor's robots are still stuck on scenario validation, and the commercial account hasn't added up yet.