By Bai Ji

Edited by Shen Xiao

"Without relying on BEV perception or rule-based code, our autonomous driving has entered the top tier of the industry, meeting the needs of users under 100,000 yuan."

This was the technical blueprint presented by Leapmotor at its 2026 Tech Day. Zhu Jiangming announced on-site that all 350,000 LiDAR-equipped models sold would receive a free upgrade to the newly released LWM World Model Intelligent Driving System, with lifetime free usage. Leapmotor's intelligent driving algorithms have now reached the top tier in the industry.

For Leapmotor, born with security genes, this is a challenge to its technical limits.

Originating from Dahua Technology and backed by security monitoring AI recognition models, Leapmotor comes with a built-in perception system. Dahua's subsidiary, Rayteer, directly supplies radars, sensors, and DMS cameras—this is Leapmotor's security legacy.

However, these security genes also limit Leapmotor's potential for intelligent driving. Security equipment focuses on single, static environments, while intelligent driving emphasizes multi-sensor coordination in complex road conditions. Under these security genes, Leapmotor remains cautious about heavy investment in algorithms.

Take the Lingxin 01 intelligent driving chip, jointly developed by Leapmotor and Dahua in 2020, as an example. It uses a 28nm process with 4.2 TOPS of computing power, while Horizon Robotics' Journey 5 from the same period uses 16nm+128 TOPS, and Tesla's FSD uses 14nm+72 TOPS.

Some car owners mentioned to "Digital Trends" that Leapmotor is best viewed as a car version of Xiaomi's Redmi—high-end hardware paired with low-end software. Those seeking advanced intelligent driving are unlikely to consider Leapmotor. "Sometimes when using Leapmotor's LCC (Lane Centering Control), the vehicle drifts to the right, requiring manual steering to the left on the highway to avoid scraping accidents."

Now, this efficiency-first approach has been extended to Leapmotor's World Model.

Leapmotor has chosen to abandon BEV representation and directly output trajectories from images to save computing power. Zhou Hongtao, Leapmotor's Senior Vice President, stated, "I don't believe that intelligent driving must require massive investment or manpower to excel."

From a sales perspective, Leapmotor resembles traditional automakers more than new energy upstarts.

In September 2026, Leapmotor delivered approximately 105,700 vehicles, more than 2.5 times that of second-place XPeng, and on par with Changan Automobile's new energy vehicles (113,200) and Geely's pure electric BEVs (125,300).

However, Leapmotor's R&D expenditure ratio has been declining during the same period, dropping from 23.6% in 2021 to just 6.1% in the first half of 2026.

In terms of gross margin, Leapmotor stood at 11.7% in the first half of 2026, lower than new energy upstarts like NIO (18.4%) and XPeng (20.6%), as well as traditional automakers like Geely (17.9%) and Changan (14.5%).

Zhu Jiangming publicly stated that Leapmotor does not pursue high gross margins even for high-end models, as this is a choice rather than an outcome. Leapmotor only aims for a 15% gross margin, with 10% saved through self-developed and vertically integrated manufacturing, leaving users to pay just 5%.

Regarding intelligent driving, Zhu believes it is necessary to enter the market early to avoid users feeling left behind once intelligent driving becomes a standard feature.

Leapmotor's transformation took a different path. Three routes were pursued simultaneously during project initiation: VLA, a two-stage spliced end-to-end system, and the World Model. After internal demo competitions, the World Model was retained due to its lowest intervention rate, while the other two were discontinued.

This differs from the approach of leading players. XPeng implemented an end-to-end architecture in May 2024, while Tesla released its end-to-end FSD V12 in August 2023. Both companies fully validated end-to-end systems before exploring World Models.

Leapmotor, however, jumped directly from a modular stage to the World Model, bypassing end-to-end systems, and shifted from manually set translation chains to understanding the physical world. Zhou Hongtao, head of Leapmotor's electronic product line, revealed that since committing to the World Model route in 2025, the code volume for intelligent driving assistance algorithms has been reduced by 90%.

The jump was achieved by bypassing the industry's mainstream BEV (Bird's Eye View) perception framework and adopting a bionic human eye perspective principle to reduce code dependency. This allows Leapmotor to minimize on-vehicle computing power requirements to as low as 200 TOPS.

In comparison, Li Auto's Mach M100 offers 1,280 TOPS per chip, while NIO's Thor exceeds 1,000 TOPS. Leapmotor's low-end computing power is just a fraction of that of leading players' flagship models.

The hidden danger (potential issue) is that BEV solves issues of overlapping fields of view and occlusion, providing coherent spatial memory for the model. Without it, the model must infer and judge within pixels, lacking a unified spatial reference in complex scenarios.

For example, multi-camera perspectives require BEV fusion to eliminate blind spots and provide a top-down world map for decision-making. Removing BEV saves computing power but forces the model to confront severe perspective distortions and physical occlusions in single-view images, requiring it to determine what, where, and which direction occluded objects are moving.

Leapmotor's solution is to add memory storage modules, with the model performing spatiotemporal alignment based on feature vectors from historical images to infer the current position and motion trends of objects.

Using the neural network's learning ability to hard-code BEV's geometric and fusion logic represents a high-risk technological innovation that substitutes algorithmic efficiency for engineering certainty. Feng Mingyue, head of Leapmotor's World Model Intelligent Driving Team, is confident:

"No technology is perfect. As long as it meets current needs after evaluation, it can be promoted and used."

In Dahua Technology's intelligent prevention and control solutions, a core selling point is intelligent compression and retrieval, achieving a video compression ratio of up to 90%, increasing retrieval channels by 10 times, and reducing costs by 50%.

This benefits from Dahua's intelligent video compression technology, which saves 50% in bandwidth and storage costs compared to H.265 while maintaining target image quality. By adding intelligent video compression equipment at network nodes, smooth retrieval and compressed playback of multiple video feeds can be achieved without altering the existing physical infrastructure.

Dahua's product logic emphasizes cost savings akin to wringing out a towel. At the hardware level, Dahua's self-developed AI chips are directly used in economical facial recognition cameras to promote market adoption of intelligent cameras. In parking lot solutions, Dahua's PoE power supply and dual-lens cameras with multi-purpose pole designs significantly reduce UTP cable deployment and cable tray costs.

Clearly, Leapmotor has inherited Dahua's cost-reduction mindset.

Not only is Leapmotor's founder, Zhu Jiangming, a co-founder of Dahua Technology, but key executives like Zhou Hongtao, Leapmotor's head of intelligent driving, also have Dahua backgrounds. Early intelligent driving R&D at Leapmotor fell under the electronic and electrical department, lacking an independent first-tier autonomous driving department until the establishment of the Intelligent Technology Research Institute in 2024.

This determines that Leapmotor's technological DNA prioritizes efficiency, focusing mainly on industrial cost reduction and hardware development, with software viewed as a secondary product. In the past, Zhu Jiangming's attitude toward intelligent driving was to follow rather than lead, avoiding excessive trial and error when the industry's prospects were uncertain.

Until 2024, according to "Lei Feng Network," Leapmotor's frontline sales discovered that intelligent driving was increasingly weighing on consumers' purchase decisions. Leapmotor was forced to enter the market, establishing the Intelligent Technology Research Institute in 2024 to fully tackle intelligent driving.

Currently, Leapmotor's Dahua genetic advantages are reaching their limits.

At the perception layer, Leapmotor has Dahua's visual algorithm team, self-developed automotive-grade AI intelligent driving chips, and a mature sensor supply chain, with autonomous technology coverage across perception, decision-making, and execution layers.

However, in May 2026, a Leapmotor C16 failed to identify a stationary obstacle during L2-level intelligent driving, only triggering braking at 9 meters away, resulting in a collision and total vehicle loss. A service manager at a Leapmotor dealership responded that intelligent driving cannot recognize stationary or irregular objects, implying no quality issues with the vehicle.

The root cause lies in the volume of Leapmotor's training data.

Leapmotor's World Model training data consists of self-collected data and real driving data uploaded by users, totaling approximately 60,000 hours during the March demo and around 200,000 hours in the current version.

In comparison, Momenta possesses 15 billion kilometers of real-world mileage, equivalent to approximately 246 million hours of driving data at 61 km/h. Huawei's ADS intelligent driving has accumulated over 16 billion kilometers of assisted driving mileage, roughly 262 million hours.

There is also a gap in data quality.

Momenta has achieved closed-loop automation for data, with the system automatically screening, cleaning, and identifying complex scenarios in the cloud, requiring only limited manual verification. Huawei possesses a full-process toolchain for data preparation, annotation, and training simulation.

In other words, excluding low-value scenarios like straight roads, sunny weather, and empty lanes, Leapmotor's data scenario coverage cannot yet match that of leading platforms, reflecting gaps in its data engineering system. Leapmotor is currently applying for multiple data annotation and quality inspection patents, with its data annotation system still under construction.

"The significance of technical analysis is limited. Whoever can sell in volume and survive until the end will win. Consumers don't care whether you use LiDAR or pure vision, VLA, World Model, or end-to-end systems," an industry analyst told "Digital Trends."

Since unveiling the World Model concept on September 16, Leapmotor's market value has surged and then retreated, with capital markets making it clear that sales feedback will ultimately determine success.

Currently, Leapmotor is making significant moves in the 100,000-yuan market. The top-spec A10, priced at 86,800 yuan, is equipped with a 128-line LiDAR and 27 perception hardware components, making it the first model under 100,000 yuan to offer end-to-end assisted driving from parking spot to parking spot. The B10 LiDAR version, priced at 119,800 yuan, supports highway and urban NOA. Meanwhile, all delivered LiDAR-equipped models are eligible for free upgrades with lifetime free usage.

The core strategy is to slash the hardware threshold for LiDAR and urban NOA to the 80,000-yuan level.

Offering high-level assisted driving permanently for free directly challenges the paid model for intelligent driving.

At the 2025 World Artificial Intelligence Conference, Wu Yongqiao, President of Bosch Intelligent Driving Control China, stated that free promotion and equal access to intelligent driving strategies must no longer be pursued. Combined assisted driving functions across all models must implement charging. Otherwise, it will bring disastrous consequences to China's intelligent assisted driving industry.

This means declining consumer willingness to pay, industry-wide Malignant involution (vicious internal competition), technological innovation giving way to price wars, resembling the death spiral of the photovoltaic industry, where industry logic shifts from who has better technology to who has greater scale and lower costs.

Currently, Huawei ADS, NIO NAD, and Tesla FSD all treat intelligent driving as a core revenue source. Leapmotor, however, relies entirely on scale to amortize the iterative costs of high-level intelligent driving, first bringing high-end hardware like LiDAR to entry-level markets to drive sales volume, then buying time to improve software.

But configuration lists do not equal intelligent driving capabilities.

Take Leapmotor's flagship model, the D19, as an example. It features dual Qualcomm Snapdragon 8797 chips with a total computing power of 1,280 TOPS, surpassing the Tesla Model 3 and Li Auto L series. However, during a 1,000-kilometer endurance challenge in April 2026, the D19's AR navigation interface suddenly went blind, with critical information like lane lines and surrounding vehicles disappearing, followed by the live comment section being closed. Leapmotor responded that the display anomaly was caused by a rendering bug, which has since been fixed.

In comparison, other automakers are narrowing Leapmotor's window of opportunity.

BYD has a larger assisted driving fleet and data scale. After selecting the Tian Shen Zhi Yan B LiDAR version for the third-generation Yuan PLUS, BYD's official website offers guarantees for parking and urban navigation. XPeng MONA has brought VLA and nationwide urban assisted driving to the 100,000-yuan level with its M03 model. Geely adopts a clearer tiered strategy for intelligent driving, with the Galaxy E5 offering L2 as standard from the Explorer version and the Starship version equipped with automatic parking.

This is Leapmotor's weak spot.

The World Model is expected to be rolled out on a large scale in 2027, with OTA updates scheduled for the first quarter to three LiDAR-equipped models. Users of models under 100,000 yuan will have to wait up to 15 months after the announcement. However, NIO NWM and XPeng's second-generation VLA have already been substantially rolled out. For Leapmotor, still in the planning stage, large-scale user testing has yet to arrive.