Author | Ni Ping Editor | Wang Ruihao
Zhu Jiangming rarely leaves his office during lunch breaks.
However, on a noon after experiencing the new version of intelligent driving, he once again drove alone, taking the D19 equipped with Intelligent Driving 4.0 for a satisfying ride of over an hour in downtown Hangzhou. In the company's internal group chat, he praised the intelligent driving team repeatedly with rare excitement.
This intelligent driving system, which the boss cannot put down, originated from an internal competition at Leapmotor lasting for several months: in October 2025, three technical routes for Intelligent Driving 2.0, 3.0, and 4.0 started simultaneously.
At the beginning of this year, in a third-party PK test, Intelligent Driving 4.0 won with a comprehensive takeover rate of only one-third of that of leading automakers—through which Leapmotor verified and clarified the technical route of the world model.
"From Intelligent Driving 2.0 to 3.0 and then to 4.0, the biggest change is the improvement in game-theoretic interaction capabilities and response speed," said Zhou Hongtao, Senior Vice President of Leapmotor.
He is the leader of this system and the soul figure of Leapmotor's technology research and development. After graduating in 2001, Zhou Hongtao joined Dahua Technology. In 2016, he followed Zhu Jiangming to join Leapmotor, promoting a series of technological developments such as full-domain in-house R&D and Electrical/Electronic Architecture (EEA).
China's intelligent driving has experienced a decade of ups and downs. After multiple switches in technical routes, it has rapidly converged, the algorithm gap has narrowed, and the first-mover advantage has diminished. Technology is no longer a scarce commodity; how to achieve large-scale implementation at an affordable cost has become a new competitive variable. At this time, it is the best opportunity for Leapmotor to catch up rapidly.
Zhou Hongtao is confident that Intelligent Driving 4.0 can achieve the goal of joining the top tier. Looking back, he believes that all decisions that reached the mass production node were correct. If he could do it all over again, "those decisions that were not good enough, just like Intelligent Driving 3.0, would be stopped in a timely manner."
From Follower to Top TierLeapmotor's Intelligent Driving 4.0 Wins the Internal Competition
For a long time, Leapmotor adopted a follower strategy in intelligent driving. It was not until 2024 that multi-path exploration rapidly converged into an end-to-end technical framework, and the technical path began to become clear; the competition in intelligent driving then shifted to the stage of commercial implementation, with technical efficiency and engineering implementation becoming key variables in the new round of competition.
The terminal market is also transforming rapidly. Frontline sales teams found in their research that, whether for winning or losing models, consumers' car purchase decisions are closely related to styling, price, safety, functional configuration, cockpit, and intelligent driving—the weight of intelligent driving is getting higher and higher.
The latest data from the Ministry of Industry and Information Technology (MIIT) shows that in 2026, the penetration rate of L2-level advanced driver assistance systems (ADAS) in new domestic vehicles has exceeded 70.5%, and the hardware installation rate of NOA (Navigate on Autopilot) high-level intelligent driving has reached 34.2%. The vast majority of mainstream new cars priced above CNY 100,000 have pre-installed a full set of intelligent driving hardware, basically achieving popularization at the hardware level.
This has directly reinforced Leapmotor's determination to increase its investment in intelligent driving. "We need to transform our past shortcomings into strengths," said Zhou Hongtao.
Internal decisions and resources have rapidly aligned in this direction. In 2024, Leapmotor integrated relevant departments and established the Intelligent Technology Research Institute, managed by Zhou Hongtao: the staff size is close to 700, computing power resources have grown by 2 times, and there are nearly 600 training servers, building a 100% in-house R&D closed loop from hardware and software to data platforms and model algorithms.
Leapmotor's Product and Technology Strategy Committee is composed of representatives from marketing, market demand, product definition, R&D, and other departments. Zhu Jiangming participates personally to make decisions on the development direction for the next five years.
In 2024, the committee set strategic goals up to 2030: to join the first tier of intelligent driving by 2026; to promote L3-level autonomous driving in 2027 and 2028, while making good technical reserves for L4-level.
Under the general direction of sprinting to the top tier, in October 2025, Leapmotor internally promoted three routes of Intelligent Driving 2.0, 3.0, and 4.0 simultaneously for technical exploration.
Among the three versions, 2.0 has been mass-produced and installed in vehicles, adopting a two-stage end-to-end approach. However, a large amount of perception information is lost during transmission, requiring rules and hand-coded programs to back up the model.
"It's like the telephone game played during team building; the person at the end receives much less information, requiring multiple confirmations. This process increases latency, so the experience and handling of various working conditions of the 2.0 model are somewhat poorer," said Zhou Hongtao.
Intelligent Driving 3.0 and 4.0 represent two upgrade directions.
3.0 continues to use the two-stage end-to-end architecture, but increases the proportion of the model and weakens the rules; 4.0 is more aggressive, focusing on the one-stage end-to-end world model—directly inputting images, supplemented by a small amount of rules to output decisions.
The two directions were hard to choose between in repeated internal discussions, which was a rare period of confusion for Leapmotor. Facing high uncertainty, Zhou Hongtao formed two teams of more than 20 people each to advance respectively.
In March this year, the two sets of algorithms underwent PK testing. The test invited third-party institutions and groups to score from subjective and objective dimensions. The results showed that the experience of Intelligent Driving 3.0 after being installed in vehicles was better than 2.0, but still could not beat 4.0.
"From Intelligent Driving 2.0 to 3.0 and then to 4.0, the biggest change is the improvement in game-theoretic interaction capabilities and response speed," said Zhou Hongtao. 2.0 needs to continuously confirm the credibility of the target, and hundreds of milliseconds may have passed; the response speed of 3.0/4.0 can be controlled within 200 milliseconds, making the physical feel smoother and the ability to handle emergencies significantly improved; 4.0 is even better than 3.0 in the anthropomorphism of the driving experience.
In horizontal comparison, 4.0 is not inferior to other automakers. Zhou Hongtao mentioned that in this same PK test, the comprehensive takeover rate of 4.0 was only one-third of that of competitors.
The comprehensive takeover rate refers to how many kilometers need to be driven before a takeover is required, including two situations: efficiency takeover and safety takeover. The former is when intelligent driving cannot bypass the vehicle ahead, and the latter is when there are violation risks such as collision or running a red light, requiring human intervention.
"We hope that the performance of intelligent driving is consistent with that of humans, without violating traffic rules or creating driving dangers, with high efficiency and smooth driving. Under this standard, the number of takeovers for Leapmotor's Intelligent Driving 4.0 is only one-third of that of leading automakers," said Zhou Hongtao. This evaluation allowed him to see the potential and upper limit of 4.0, clarifying the subsequent technical direction. Since then, Leapmotor has concentrated its R&D resources on Intelligent Driving 4.0.
As the Intelligent Driving Industry Rolls Towards Large Computing PowerLeapmotor Chooses a More Minimalist Path
While upgrading from two-stage to one-stage end-to-end, Leapmotor has built a set of world models for Intelligent Driving 4.0 on both the cloud and the vehicle end.
The parameters of the cloud base model are about 128B, which is the top configuration in the industry. A larger parameter scale brings stronger memory for complex scenarios and corner case handling capabilities.
Through knowledge distillation and network structure optimization, this ultra-large model is compressed into a lightweight version and deployed to the vehicle end, where it can run smoothly on a 200 TOPS computing power platform.
Even after extreme compression, the vehicle-end model still retains core capabilities such as interaction prediction, enabling real-time environmental perception and decision-making while driving. This architecture of "heavy training on the cloud and light inference on the vehicle end" allows Leapmotor to achieve effective implementation of the world model without relying on large computing power.
Zhu Jiangming once frankly expressed at the launch of the A10: if it is purely for driving, the computing power of a chip similar to 8650 is enough to make it basically usable. In other words, 200 TOPS of computing power is sufficient to meet the needs of intelligent driving.
Zhou Hongtao added in the exclusive interview that the response time of Leapmotor's world model on the 200 TOPS platform is about 200 milliseconds—faster than the time it takes for a human to see and make a hand reaction.
When the industry generally superimposes multi-modal information such as images, text, and BEV (Bird's Eye View) in the one-stage end-to-end approach, Leapmotor's R&D team chose a more minimalist path. Their judgment is based on a simple observation: humans do not construct a BEV bird's-eye view in their brains while driving. Therefore, the solution for Intelligent Driving 4.0 is determined to be pure image input directly outputting trajectories, without redundant conversion.
The model itself does not directly output control signals, but only generates the expected driving trajectory within the next 4 seconds. Therefore, the R&D team connects an extremely small amount of rule layer on this basis to translate the trajectory intention into throttle and steering wheel commands.
In this process, post-processing must be sensitive and closely aligned with the model's intention, rather than following the rule experience from the past two-stage era.
"The team has reached a consensus internally: trust the model and assist the model," said Zhou Hongtao. However, this technical leader also maintains caution and clarity: no matter how many 9s the model accuracy reaches, the 1 representing unknown scenarios always exists, and ultimately, a small amount of rules must be used as a backup to ensure the safety boundary.
The training data for this set of world models comes from the accumulation of Leapmotor's past vehicle models. The BC models are smaller in size, and their LiDAR data is used as Ground Truth to supervise the pre-training process of the model; the D series models are larger, and the sensor installation positions differ from those of the BC models, which also provides diverse sample inputs for the model.
After cleaning and screening, about 30 million video clips enter the pre-training stage of the 4.0 model. "The richer the samples the model has seen, the easier it is for a single set of models to directly generalize to different vehicle models, and the lower the adaptation cost," said Zhou Hongtao.
In addition to requiring sample diversification, the team also screens out repetitive and non-standard driving behaviors—Leapmotor once trained a version of the model in its early days where the vehicle would inexplicably stop in place when encountering a green light. It was only after cleaning the data later that they discovered the reason: in some samples collected at night, the driver was playing with their phone while waiting for the light. Therefore, Leapmotor re-established data collection rules and reward/punishment mechanisms to improve data quality from the source.
In addition to data provided by its own fleet, Leapmotor also supplements collection by outsourcing to high-quality drivers. Currently, Zhou Hongtao and his team are promoting generalization testing in various cities across the country, iterating and optimizing rapidly.
Automated Driving Technology Drives Architecture Innovation in Reverse
Looking back, Leapmotor wrote the first line of code for the automated driving 4.0 version at the end of 2025, and achieved phased results in the PK test in March this year. In just a few months, Leapmotor completed the key leap from project initiation to finalization.
Supporting this speed is the synchronous evolution of automated driving algorithms and Electrical/Electronic Architecture (EEA).
Over the past 8 years, Leapmotor's Four-Leaf Clover architecture has iterated a generation every two years. The team continuously observes industry development trends, finds current bottlenecks and pain points, and continuously improves issues such as computing power distribution and communication efficiency in the next-generation architecture.
The latest EEA4.0 central integrated architecture achieves multi-domain fusion, central computing, and zonal control. The proportion of Ethernet has been significantly increased, providing a foundational support for the development of Automated Driving.
The cost reduction effect brought by architecture integration has also opened up commercial space for the downward penetration of automated driving.
Zhou Hongtao revealed that the cost of Leapmotor's 4.0 architecture is only 40% of that in the 1.0 era. The significant decrease in underlying costs has freed up budget for BOM expenditures on LiDAR, cameras, and automated driving chips, and has also become the underlying commercial premise for Leapmotor to push its self-developed world model 4.0 to A-platform vehicle models priced around CNY 100,000.
And the further iteration from EEA4.0 to EEA5.0 also means that the demand for automated driving technology has begun to drive architecture innovation in reverse.
The evolution logic of the EEA5.0 architecture is central computing combined with simple control in edge zones. By directly connecting various controllers and actuators through Ethernet, it increases transmission bandwidth, improves communication efficiency, and enhances interactive experience.
This forms a high degree of isomorphism with the world model + one-stage end-to-end route adopted by automated driving 4.0—moving towards integrated transformation, reducing information transmission loss, and forming a mutually responsive and synchronously evolving relationship between architecture iteration and automated driving technology.
In the process of architecture iteration, massive data computing and continuous cross-domain task execution also put forward higher requirements for power and electrical architecture.
The traditional 12V system is approaching its usage limit, and high-power output is often accompanied by problems such as rising costs, excessive energy loss, and fast battery drain. Therefore, Leapmotor's EEA5.0 architecture proposes a full-domain 48V system, which can reduce weight and heat loss, and improve energy efficiency.
However, the supply chain ecosystem is not yet mature. Leapmotor needs to wait patiently and also needs to persuade suppliers to transform together. "The iteration from EEA4.0 to EEA5.0 may take longer, because we have already reached the uncharted territory," said Zhou Hongtao.
The continuous advancement of technology corresponds to a broader business map for Leapmotor.
The technical architecture of central brain plus zonal control, and the world model technical route, are not only the foundation for vehicle intelligence, but also paving the way for embodied intelligence.
The new business is already reflected in the organizational structure. Currently, the Intelligent Technology Research Institute has five departments: data closed loop, end-to-end model, mass production development, computing power platform, and embodied intelligence.
Zhou Hongtao revealed that the robot prototype has been developed, but Leapmotor does not plan to make performance-oriented products, nor does it consider general scenarios such as supermarket shopping guides and domestic services.
"Our robots can dance and run without any problem now, but they cannot yet create practical value. Now is not the time to release them," he said. The short-term goal of embodied intelligence is very clear: anchoring on the two major scenarios of factory automation and logistics transportation, and reversely strengthening the cost competitiveness of the whole vehicle by improving the automation efficiency of production lines and warehousing.
In terms of product form, there are technical reserves in directions such as humanoid and wheeled, and ultimately, the most suitable product form will be matched according to the target application scenarios.
The technical route of embodied intelligence has not yet converged, and the commercial prospects are not yet clear. Measuring the investment intensity of this business is closely related to how much it can empower the main vehicle manufacturing business.
This is also Leapmotor's consistent decision-making logic.
It does not pursue absolute leadership in technological iteration, is not obsessed with large computing power, and does not tell sexy stories under the technological paradigm. Just like the example Zhu Jiangming often cites: if you put photos taken by four mobile phones of different brands in front of people, it is hard to tell which phone took which photo—technology will eventually converge, and automated driving is no exception.
The real winning hand lies elsewhere. Leapmotor firmly believes that at the right time, pushing mature technology to a sufficiently large market to meet the needs of the majority can bring about qualitative change.
The massive installed base will in turn feed back data and drive the next round of iteration—market expansion and technology popularization, cycling back and forth like this.
From June to August this year, Leapmotor's monthly sales exceeded 100,000 units for three consecutive months, gathering strength to sprint towards a million sales by the end of the year. Automated driving 4.0 will become an important part of the new round of expansion.
This automaker, which has always won with endurance, efficiency, and scale, has once again waited for its own time.