In September, Zhuoyue Technology released “Chengxing Intelligent Driving 2.5.” After various computing power solutions, Zhuoyue brought the “two-stage” end-to-end system to a relatively mature phase. Zhuoyue may be the only company in the industry capable of implementing urban NOA based on a 7V+32Tops chip. Most competitors deploy urban NOA solutions based on 254Tops or even higher computing power. Now, “hardware agnostic” is no longer Zhuoyue Technology’s only competitive edge. Even in the “unrestricted combat” that does not limit hardware costs, Zhuoyue’s actual road test performance surprises many. Which is better, “two-stage” or “one-stage”? Everyone knows Zhuoyue Technology insists on “two-stage” end-to-end. Since its establishment in 2016, Zhuoyue has navigated various stages, from CV small models and regulatory control routes to BEV/occ and end-to-end, VLA, and world models. Each step has built trust in its technical capabilities. Technology must inherit from previous advancements; sudden abilities that appear out of nowhere often lack substance. Before 2023, suppliers and OEMs in intelligent driving had not reached a consensus on the technical path. High-precision mapping plus regulatory control competed evenly with light mapping plus BEV/occ. By 2023, end-to-end solutions began to dominate. People divided end-to-end into two types: “two-stage” and “one-stage.” Both lack precise definitions. “One-stage” likely refers to a process from perception to execution, using a single model with internal data flow as a “black box.” The difference with “two-stage” is that it separates perception and regulatory control into two distinct models, embedding an artificial data interface in between. The theoretical advantage of “one-stage” lies in minimal information loss, suggesting a higher potential. However, its downside is poor interpretability, leading to a lower theoretical limit. The benefit of “two-stage” is observable interface data, making perception output a “white box.” This reduces the unpredictability of errors in intelligent driving.
All manufacturers set bottom-line rules, whether “two-stage” or “one-stage.” The “one-stage” design is inherently more challenging than the “two-stage.” It has a longer tuning loop and greater system uncertainty. In reality, both “one-stage” and “two-stage” have variations, blurring their boundaries. The end-to-end parameter count matches that of medium models and runs on Nvidia Orin-X. As the next-generation Thor faces delays, the first batch of simplified Thor-U models began deployment in May. Everyone aims to increase the parameter count of transfer models to tackle long-tail issues. However, excessive parameters and lengthy decision-making chains can slow execution. Thus, the design incorporates both fast and slow thinking chains. Zhuoyu operates in a “sparsely populated” niche. Can its 32 Tops of computing power handle complex urban traffic? The capability of transfer models largely depends on the underlying large model training system. Zhuoyu’s core innovation builds on the “two-stage” end-to-end architecture. It employs interactive modeling and reinforcement learning to overcome the limitations of the classic “two-stage.”
Interactive modeling technology couples dynamic and static environmental information with navigation data. This multimodal integration makes lane changing more natural and human-like. This innovation addresses the decision delay caused by the separation of perception and control in the “two-stage” process. In tests, Zhuoyue’s 7V computing system achieved a lane change success rate exceeding 90% in congested areas, matching skilled human drivers. The system enhances recognition and perception of long-tail scenarios, such as “phantom vehicles,” by predicting dynamic targets and integrating them with the vehicle’s planning. This approach allows for proactive braking and evasive maneuvers. Many suppliers focus on reinforcement learning. Zhuoyue employs heavy simulation combined with reinforcement learning. The model’s ability to identify atypical scenarios improves significantly. Zhuoyue has built a database of over one million clips for in-depth data mining on more than ten safety scenarios, enhancing trajectory safety. In complex urban road tests, success rates for unprotected left turns and narrow road encounters improved significantly compared to the classic “two-stage” method. The key is that this optimization does not require more decision-making time. With this large model training foundation, Zhuoyue optimizes algorithm efficiency through precise scheduling of computing resources and memory usage on a mid-level computing platform. This also reflects that most end-to-end models still have significant potential for algorithm efficiency. Open test data shows that on a 32 Tops computing platform, Zhuoyu’s “Chenghang Smart Driving” maintains a perception frame rate above 25 FPS. It also keeps planning and control delays under 150 ms, meeting the demands for speed and depth in decision-making. With this technological foundation, Zhuoyu’s system can fit into entry-level products priced between 100,000 and 150,000 yuan (21060$). Many companies talk about “smart driving equality,” but Zhuoyu can help its OEM clients fulfill that promise. The large model training and推 support behind the “soft and hard integration” shines through. Zhuoyu’s capabilities extend beyond algorithms. It offers a comprehensive solution that integrates hardware and software. Zhuoyu built its core perception solution on an inertial navigation stereo vision system. This system mimics human binocular vision, using parallax to obtain depth information and generate dense visual point clouds. It can detect any obstacles and achieve precise avoidance, even in low visibility conditions like nighttime or rainy weather. Zhuoyu’s two-stage end-to-end model integrates information from multiple sources. It uses a safety reasoning framework to select the optimal path. This design retains the advantages of end-to-end systems while providing necessary safety redundancy.
Zhuoyu did not simply use cost-effectiveness as its trump card on high-performance platforms. Zhuoyu solved the conflicts between camera and lidar data. The VLA large model, developed based on the NVIDIA Thor chip, is about to enter mass production. Zhuoyu released the high-perception end-to-end 2.0 model. This includes the end-to-end model and the VLA model. Some world models based on VLA can achieve seamless parking and personalized driving experiences. Users can also control functions via voice. Zhuoyu built a smart computing center and big data platform on Alibaba Cloud. This serves as a training and promotion platform. It supports end-to-end models, achieving rapid iteration and data feedback capabilities. Many manufacturers have yet to recognize this core capability, but Zhuoyu has already established it. The business route has expanded. In 2025, Zhuoyu announced its European headquarters in Brunswick, Germany, during the Munich Auto Show. This move lays the foundation for its technology to enter the European market. Although the complete vehicle brands it partners with must pass all EU standards, Zhuoyu can handle inspections independently, relieving manufacturers of concern. The IQ.PILOT system, co-developed with Volkswagen, passed the A-SPICE CL2 certification. It became the first intelligent driving supplier in China to receive joint approval from Volkswagen Group, FAW-Volkswagen, and SAIC Volkswagen. Currently, Zhuoyu has partnerships with 10 manufacturers and over a dozen brand clients. This year, Zhuoyu’s business has climbed to new heights. In October, it widely promoted the medium-computing power city NOA. In August, it mass-produced the high-speed NOA solution “Oil-Electric Smart Driving Equality” for fuel vehicles. This fulfilled a long-held dream for many brands with extensive fuel vehicle lineups. Zhuoyu became the first to offer a commercialized solution, gaining recognition from major clients like Volkswagen, Audi, Jietu, and Chery. This marks a milestone achievement for Zhuoyu in 2023, following its spin-off from DJI. This success even overshadowed news of FAW becoming its largest single shareholder. Zhuoyu Technology showcased a “two-stage” end-to-end architecture, from technical path selection to market application validation. It demonstrated the journey from principles to commercial applications. Under current technical conditions, it implemented various supply solutions across multiple hardware platforms. The value and commercial viability are proportional. In the past two years, we have witnessed smart driving technology transition from labs to large-scale commercial use. Zhuoyu Technology proved that a medium-computing power platform combined with optimized algorithms can deliver a reliable and practical smart driving experience. In a competitive landscape where OEMs and smart driving suppliers vie for users, Zhuoyu has secured capital ties with major clients. Even as an independent supplier, its unique reliability, economic efficiency, and scalability have earned votes from an increasing number of B-end clients. With undeniable momentum, Zhuoyu Technology secures its position as a Tier 1 leader in intelligent driving.