현재, 도시 NOA(인텔리전트 드라이빙) 기능은 15만 위안 클래스의 차량 모델에까지 보급되어, 경쟁 장벽이 소실된 것으로 업계의 균질화가 가속하고 있습니다. 이 단계에서 L3은 자동차 제조업체가 사용자를 얻고 브랜드 가치를 높이는 중요한 획기적인 역할을합니다. L3에서 브레이크 스루를 완수함으로써, 제조업체는 사용자가 요구하는 '보다 안심하고 안전한' 운전이라는 고차원의 요구에 부응하여 차별화된 우위성을 확립할 수 있습니다. L3은 기술력의 시금석인 동시에 브랜드 가치를 증폭시키는 역할도 담당하고 있습니다. 진정한 인텔리전트 드라이빙을 위한 중요한 단계인 L3은 규제적합과 첨단 센서 퓨전과 같은 과제를 극복할 필요가 있으며, 그 신뢰성은 제조업체의 기술력을 직접 반영합니다. L3의 양산을 선행하는 제조업체는 「기술적 리더십」이나 「하이엔드 인텔리전스」의 브랜드 이미지를 신속하게 확립해, 전모델의 가치 향상을 촉진하는 것과 동시에, 경쟁사와의 차이를 넓힐 수 있습니다.
Geely Zeekr의 경우, 인텔리전트 드라이빙의 진화 경로는 명확하게 L3을 지향합니다. 2023년 12월에 자사 개발의 풀 스택 인텔리전트 드라이빙 시스템을 발표해 고속도로 NOA와 APA를 실현했습니다. 2024년 12월에는 완전한 지도 프리 도시 NOA를 전개했습니다. 2026년 6월에는 Door-to-Door(D2D) 기능을 탑재할 예정입니다. L3과 레벨 4에서의 혁신은 미래의 기술적 진화에서 핵심적인 방향이 될 것입니다.
동향 1 : 소비자에 의한 고레벨 인텔리전트 드라이빙 기술에 대한 수요가 시장 구조를 재구축하고 있으며, 33%의 소비자가 도시 NOA에서 L3/4 기능으로의 업그레이드를 희망하고 있습니다.
신발매차의 데이터에 의하면, 2023-2025년에 중국의 승용차에 있어서 인텔리전트 드라이빙 탑재수는 양극화가 현저했습니다. L2.5/L2.9의 고레벨 인텔리전트 드라이빙 기술이 비약적으로 성장하는 한편, 종래의 L1-L2 기능은 감소 경향이 계속되어 업계의 인텔리전트화와 반복 가속의 조류가 명확하게 나타났습니다. 2023년 시점에서 L2.5/L2.9의 인텔리전트 드라이빙 기술은 시장에서 여전히 틈새 시장이었으며, 신발매 모델에 대한 탑재율은 각각 불과 4.57%, 3.3%에 그쳤습니다. 그러나 모두 2025년 1-4월에 급성장했습니다. L2.5 인텔리전트 드라이빙을 탑재한 신차의 비율은 34.8%로 급상승, L2.9를 탑재한 신차의 비율은 34.82%를 차지해 매우 높은 시장 침투율을 나타냈습니다. 이와는 대조적으로, 종래의 L1-L2 레벨 인텔리전트 드라이빙 기술의 탑재율은 동기간에 감소 경향을 보였습니다. 이 양극화의 동향은 소비자의 첨단 인텔리전트 드라이빙 기술에 대한 선호가 시장 공급 구조를 바꾸고 있음을 명확하게 보여주며, 높은 수준의 인텔리전트 드라이빙 기술이 신차 시장에서 경쟁의 핵심 초점이 되고 있습니다.
이 동향의 핵심 촉진요인은 도시 NOA와 고속도로 NOA 기능을 탑재하는 가속에 있습니다. 이 두 가지 기능은 인텔리전트 드라이빙 기술이 '개념'에서 '실용화'로 전환하기 위한 중요한 담당자일 뿐만 아니라 '소비자 교육'이라는 중요한 역할도 담당하고 있습니다. 통근이나 장거리 이동 등의 고주파 이용 장면에서 소비자는 고차원 인텔리전트 드라이빙의 가치를 직감적으로 실감하고 기술에 대한 인지와 신뢰를 서서히 구축하고 있습니다. 「실용화→만족감→상위 기능에의 욕구」라는 수요의 순환 구조가, 유저의 선진의 인텔리전트 드라이빙에 대한 기대를 한층 더 가속시키고 있습니다.
또한 사용자는 오랫동안 ADAS와 "공존"하는 동안 기존의 인텔리전트 드라이빙 기술의 한계를 이해하고 기본적인 신뢰를 구축해 왔습니다. 그러나 L1-L2 레벨의 인텔리전트 드라이빙 기술은 복잡한 도시도로 상황에 대응이 곤란하다는 기능적인 한계에 가까워지고 있으며, 이 한계가 사용자 수요를 보다 선진적인 인텔리전트 드라이빙 기술로 이행시키고 있습니다.
McKinsey의 China Auto Consumer Insights 2025의 조사 데이터는 '기술 보급 → 소비자 인지'라는 선순환을 직접 뒷받침합니다. 2023년에 비해 인텔리전트 드라이빙 기술에 대한 소비자의 수용도와 만족도는 현저하게 향상되고 있습니다. 특히 핵심 기능에 있어서는 2024년 기준에서 현행의 도시 NOA(네비게이션 첨부 인텔리전트 드라이빙)에 만족하고 있는 유저는 46%에 이르렀고, 도시 NOA 유저의 33%가 기존의 도시 NOA를 L3/L4 레벨로 업그레이드하는 것을 명확하게 희망하고 있습니다.
이 보고서는 중국 자동차 산업에 대해 조사했으며, 15개 OEM(중국 제조업체 8개, 외국 제조업체 7개 포함)과 9개 회사의 Tier 1 공급업체(반도체, LiDAR, 도메인 컨트롤러, ADAS 등을 다루는)에 대한 상세한 조사를 통해 L3 인텔리전트 드라이빙의 핵심 레이아웃을 분석했습니다.
L3 Research: The Window of Opportunity Has Arrived - Eight Trends in L3 Layout of OEMs and Tier 1 Suppliers
Through in-depth research on 15 OEMs (including 8 Chinese and 7 foreign OEMs) and 9 Tier 1 suppliers (covering chips, lidar, domain controllers, ADAS, etc.), ResearchInChina analyzes the core layout of L3 intelligent driving of the two groups. For OEMs, this report comprehensively combs through their L3 intelligent vehicle development strategies, key launch nodes, and first L3 models, as well as sensor hardware solutions, intelligent driving chip selection, technology path planning, and redundancy strategy design. For Tier 1 suppliers, it focuses on exploring the R&D and implementation progress of their L3 intelligent driving products. Based on the above research, it finally summarizes eight major development trends of L3 intelligent driving in the Chinese market over the next 3 years.
Currently, urban NOA has been extended to vehicle models priced at RMB150,000. The competitive barrier disappears and industry homogenization intensifies. At this time, L3 has become a key breakthrough for OEMs to compete for users and achieve brand upgrading. Only by making breakthroughs in L3 can OEMs meet users' high-level demands for "more worry-free and safer" driving, and establish differentiated advantages. L3 is not only a touchstone for technical strength but also an amplifier of brand value. As a crucial step towards true autonomous driving, L3 needs to overcome challenges such as regulatory compliance and advanced sensor fusion, and its reliability directly reflects OEMs' technical capabilities. OEMs that take the lead in mass-producing L3 can quickly establish labels of "technological leadership" and "high-end intelligence", drive up the value of the full range of their models, and widen the gap with competitors.
In Geely Zeekr's case, its intelligent driving evolution path clearly points to L3: launched a self-developed full-stack intelligent driving system in December 2023, realizing highway NOA and APA; fully rolled out mapfree urban NOA in December 2024; will implement Door-to-Door (D2D) function in June 2026. Making breakthroughs in L3 and L4 is the core direction of its next technical evolution.
Trend 1: Consumers' Demand for Higher-level Intelligent Driving Functions Is Reshaping the Market Structure, with 33% of Consumers Hoping to Upgrade Urban NOA to L3/L4 Functions.
From the data of newly launched vehicles, the installation of intelligent driving in passenger cars in China featured a greatly polarized pattern from 2023 to 2025: L2.5/L2.9 high-level intelligent driving functions enjoyed leapfrog growth, while traditional L1-L2+ intelligent driving functions continued to decline, clearly reflecting the trend of faster industry intelligence and iteration. In 2023, L2.5 and L2.9 intelligent driving were still niche configurations in the market, with installation rates of only 4.57% and 3.3% in newly launched models, respectively. However, from January to April 2025, both boomed: the proportion of new cars equipped with L2.5 intelligent driving soared to 34.8%, and those with L2.9 even took a 34.82% share, showing a very high market penetration. In sharp contrast, the installation rate of traditional L1-L2+ intelligent driving functions was on the decline during this period. This polarization trend clearly indicates that consumers' preference for higher-level intelligent driving functions has begun to reshape the market supply structure, and high-level intelligent driving is gradually becoming the core focus of competition in the new car market.
The core driving force behind this trend lies in faster implementation of urban NOA and highway NOA functions. These two types of functions are not only key carriers for autonomous driving technology to move from "concept" to "practical application" but also assume the important role of "consumer education". In high-frequency scenarios such as daily commuting and long-distance driving, they allow consumers to intuitively perceive the value of high-level intelligent driving and gradually establish cognition and trust in the technology. The demand closed loop of "practical use - satisfaction - desire for upgrading" further catalyzes users' expectations for higher-level intelligent driving.
Moreover, as users have long "coexisted" with ADAS, they have gradually figured out the capability boundaries of existing intelligent driving and have built up basic trust. However, the functional ceiling of L1-L2 intelligent driving, e.g., difficult to cope with complex urban road conditions, is about to be reached, and this limitation is continuously encouraging user demand to shift to higher-level intelligent driving.
The research data from the China Auto Consumer Insights 2025 by McKinsey directly confirms the positive cycle of "technology penetration - consumer recognition". Compared with 2023, consumers' acceptance and satisfaction with autonomous driving functions have significantly improved. Specifically for core functions, 46% of users were satisfied with the current urban NOA in 2024, and 33% of urban NOA users clearly hoped to upgrade the existing urban NOA to L3/L4.
Trend 2: From the Supply Side, the Window of Opportunity Brought by L3 Is Clear, Multiple Chinese OEMs Have Taken the Period from 2025 to 2027 as a Critical Phase for Mass Production and Installation of L3 Intelligent Driving, and Pre-embedded Hardware Becomes the Mainstream Strategy.
From the supply side, the window of opportunity for commercialization of L3 intelligent driving is clear. Leading OEMs such as NIO, Xpeng, Geely, and Huawei-affiliated OEMs have regarded 2025-2026 as the critical period for mass production. Pre-embedded hardware has become the mainstream industry strategy: by pre-equipping components such as lidar and high-compute chips, they can quickly activate functions to win a place and gain first-mover advantages after regulatory relaxation.
Their commercial implementation follows a clear path of "highway -> urban area", "closed -> open", and "business -> consumer". Policy breakthroughs and cost reduction constitute a dual engine: in 2025, Beijing and Shanghai have explicitly defined the liability division for highway L3 accidents, and highway L3 functions of players such as Huawei and Xpeng have been delivered for production vehicles; the significant reduction in cost of hardware such as lidar has paved the way for technology popularization. Wherein, highway scenarios have become the first "test field" to implement L3 due to highly structured roads and easy unification of regulations.
Although the consumer market still needs to break through the bottlenecks of user trust and cost sensitivity, as the industry chain matures, it is expected that L3 models will enter the mid-range price segment in the next 3 years. The concentrated mass production of multiple OEMs during 2025-2026 indicate that L3 technology has entered a new phase of "large-scale commercial implementation" from "testing and verification".
Trend 3: L3 and L4 Dual-line Layout: OEMs' Technical Collaboration and Ecosystem Competition
Some leading OEMs are betting on both L3 and L4, which is essentially a deep binding at the technology, capital, and strategy levels. By means of two-way technical enablement and commercial complementarity, they build an irreproducible competitive barrier.
At the technology level, L3 and L4 form a "symbiotic evolution" closed loop. Both are highly universal in hardware such as lidar and high-level intelligent driving chips, as well as in automotive redundancy design, with interoperable core capabilities. L3 production vehicles can collect a mass of edge case data such as "takeover" scenarios, becoming a "training library" for L4 algorithms. The high-level algorithms of L4, after being downscaled to designated scenarios, can directly improve the performance reliability of L3. This synergy of "data feedback + technology downscaling" enables the two technology paths to achieve an iteration efficiency of 1+1>2.
At the commercial level, both form an ecosystem combination of "short-term blood transfusion + long-term occupation". L3 quickly recovers funds and verifies the market through private car sales, transfusing L4 R&D. L4 targets the Robotaxi market worth RMB1 trillion and lays out the future mobility ecosystem. More importantly, private cars and Robotaxi fleets can share resources such as HD maps and cloud platforms to form operational synergy. OEMs that take the lead in overcoming L4 are expected to become the definers of the future mobility ecosystem. This dual-line strategy is not only a pragmatic choice to reduce technical R&D risks but also a strategic layout to have a say in the intelligent driving era.
Trend 4: L3 Technology Path Shows a "Three-legged Stool" Pattern: Independent R&D, Dual-track (Co-development + Independent R&D), and External Suppliers.
In China, L3 intelligent driving technology path has formed a "three-legged stool" pattern of "full-stack independent R&D, co-development + independent R&D, and external cooperation". This is essentially a differentiated choice of OEMs based on their technical reserves, capital strength, and strategic rhythm - seeking the optimal balance between "technical sovereignty" and "commercial efficiency".
Full-stack Independent R&D: Exchange high investment for long-term technical moat
Leading OEMs such as NIO, Xpeng, Li Auto, and Geely have anchored full-stack independent R&D, the core of which is to master the full-link dominance of underlying hardware (such as self-developed chip adaptation) and top-layer algorithms (end-to-end large models). This model can build an exclusive data closed loop, continuously collect edge data such as "takeover" scenarios via production vehicles to reversely feed algorithm iteration, and build an irreproducible technical barrier. However, the cost is high, and there are technical trial and error and cycle risks.
Co-development + Independent R&D: Balance independent control and R&D efficiency
Cases such as SAIC IM's co-development with Momenta and Dongfeng Voyah's "strategy implementation by brand" (independent R&D for Taishan + Dreamer equipped with Huawei's solution) represent the flexibility of the mixed route. Its core logic is "independent R&D or co-development of core technologies + outsourcing of the non-core": OEMs control key links such as decision algorithms, and entrust heavy-asset links such as perception fusion and data annotation to professional partners. This not only avoids resource waste of fully independent R&D but also gets rid of the risk of depending on single ones. This model has become the preferred choice for most traditional OEMs. For example, BYD, while independently developing "God's Eye", works with Momenta to implement high-level functions.
Choosing External Suppliers: Use mature solutions to take a place quickly and shorten L3 R&D cycle
Typified by Huawei's cooperation with OEMs, this model quickly enters the market by virtue of "packaged solutions". With its ADS 4.0 system featuring integration of "chip - algorithm - redundancy architecture", Huawei covers more than 7 OEMs. It empowers JAC STELATO S800 to implement highway L3 intelligent driving and build a flagship intelligent model.
Trend 5: Multi-channel Lidar Becomes an Important Choice for OEMs to Lay out L3 Intelligent Driving and Ensure Safety Redundancy.
Global L3 intelligent driving sensor solutions show a clear differentiation: only Tesla and Xpeng adhere to the vision-only route, while other mainstream OEMs inside and outside China and pilot manufacturers take lidar as the core configuration. Chinese ones include Huawei-affiliated OEMs, Geely, GAC, SAIC IM, and NIO; foreign ones such as BMW, Mercedes-Benz, Honda that have piloted L3, and European and American giants that have not conducted road tests.
Markus Schafer, CTO of Mercedes-Benz, pointed out that L3 requires multi-sensor redundancy to ensure safety, and as vehicle speed increases, higher-performance lidar is needed for long-distance small obstacle detection, reserving sufficient processing time for the system and the driver. Chen Xiaozhi, Chief AI Technology Officer of Zhuoyu Technology, also emphasized that the hardware safety redundancy of L3 requires sensor complementarity (not just relying on algorithms), and the core value of lidar is to provide safety redundancy.
As the core indicator of lidar resolution (representing the number of vertical laser beams), the number of channels directly matches the upgrade of intelligent driving levels: early 16-channel lidars are suitable for low-speed scenarios, the 32/64-channel serve low-to-mid level ADAS, and currently the 128-channel have become a mainstream automotive solution. L3 models require >=128-channel lidars, and the mainstream configuration has been upgraded to 192-channel, 520-channel, or even 700-channel. The dense point cloud brought by multi-channel lidars can realize accurate recognition of small obstacles 170 meters away, which is an essential safety requirement for L3 in the scenario of liability transfer.
Trend 6: The Computing Power Required for L3 Intelligent Driving Shows an Exponential Leap, with 1000TOPS Becoming the Mainstream Threshold.
The computing power required for autonomous driving is not blindly piled up, but is deeply bound to levels, scenarios, and algorithm models. L2 deals with basic scenarios such as lane keeping and adaptive cruise control, and >=50TOPS dense computing power is sufficient. Excessive stacking will only lead to resource waste and high cost.
Due to the need to take on the main driving responsibility, L3 needs to cope with complex urban traffic, various traffic participants, dynamic environmental changes and other scenarios. It requires large-scale neural network models for real-time reasoning. The expansion of end-to-end large model parameters results in the demand for higher vehicle computing power. After combining end-to-end technology and VLM into VLA, the vehicle-side model parameters become larger. It not only needs efficient real-time reasoning capability but also has the ability to recognize the complex world and give suggestions. Deploying VLA models will pose quite high requirements for vehicle chip hardware. The demand for sparse computing power directly jumps to 1000-2000TOPS level, and the dense computing power threshold rises to >=200TOPS.
The sparse acceleration ratio varies in scenarios. For structured roads such as highways, the proportion of effective information exceeds 50%, the sparsity is low, the acceleration ratio is only 2-3 times, and the equivalent computing power of 400-600TOPS is sufficient. For long-tail scenarios of complex urban road conditions, the proportion of effective information is lower than 10%, the sparsity is high, the acceleration ratio can reach 8-10 times, and the equivalent computing power can be increased to 1600-2000TOPS, which accurately matches the computing power requirements of complex environments.
Trend 7: Device-cloud Collaboration, the Core Architecture for Breaking through Computing Power Constraints in L3 intelligent Driving Layout
"Device-cloud collaboration" has become a consensual choice for mainstream OEMs laying out L3 intelligent driving to break through computing power. Its essence is to solve the core contradiction between the performance requirements of large models and the limitations of vehicle computing power through the division of labor of building capabilities in the cloud and implementing applications on vehicles. Xpeng's technical practice is a typical example of this path.
The underlying technical logic is driven by the Scaling Law: the number of parameters and data volume directly determine the performance of models, but vehicle computing power is difficult to support the operation of 10-billion-parameter large models. By training models with 1 billion to 72 billion parameters and feeding more than 20 million clips of video data, Xpeng's team first verified that this law is still available in an autonomous driving VLA model. The 72-billion-parameter cloud large model can accurately handle complex scenarios, and then generate a small model suitable for vehicles using distillation technology, which can preserve core capabilities to the greatest extent possible and break through computing power constraints.
The "Xpeng World Foundation Model" released in April 2025 is the carrier of implementing this logic. As a cross-terminal "super parent body", it realizes full-link production through the "cloud model factory" built by Xpeng: forming a closed loop from multi-modal pre-training, reinforcement learning post-training, to model distillation and vehicle-end deployment. Relying on the 10,000-card intelligent computing cluster, the iteration cycle is compressed to an average of once every 5 days.
Its evolution core lies in the "Dual Loop Collaboration Mechanism": the Inner Loop completes the efficient transfer of large model capabilities to the vehicle-end in three stages of "pre-training - reinforcement learning - distillation"; the Outer Loop continuously reversely feeds cloud model iteration relying on the perception data of real vehicles, user feedback and extreme cases, completely solving the problem of disconnection between simulation and real scenarios. This closed loop of "training intelligence in the cloud, using intelligence at the vehicle-end, and returning intelligence with data" not only enables a small vehicle-end model to have generalization capabilities close to that of large models but also realizes continuous self-evolution of intelligent driving systems, laying a technical foundation for safe implementation of L3 and advancement to L4.
Trend 8: L3 Intelligent Driving Moves Towards End-to-end 2.0, and VLA Becomes One of the Mainstream Routes to Break Through Experience Bottlenecks for L3 Intelligent Driving.
L3 intelligent driving is moving from "modular splicing" to the "end-to-end 2.0" era. The core evolution logic is the deepening of multi-modal fusion. The combination of "VLA (Vision-Language-Action model) + device-cloud collaboration + world model" is becoming the mainstream path to break through technical limitations and realize commercial use. SAIC IM's three-stage evolution route accurately embodies this leap process from "technology availability" to "experience reliability".
2025 is the "foundation building period": by implementing the one-model end-to-end (E2E) architecture, break the module barriers of traditional perception, decision, and control, and realize lossless information transmission and global optimization, with L3 having been 90% production-ready in terms of technical maturity. This step solves the core sore points of "information loss and accumulated errors" in modular systems and lays a solid foundation for high-level intelligent driving.
2026 enters the "capability leap period": introducing multi-modal large models (E2E+VLM) on the basis of end-to-end enables the system to have initial scenario semantic understanding capabilities, integrating visual perception, voice commands, and map information to make decisions, instead of relying solely on sensor data. This upgrade directly makes up for key shortcomings in L3 commercialization, making it meet the mass production condition of "high reliability".
2027 and beyond move towards the "ultimate form period": evolve into full-link multi-modal end-to-end (VLA), and realize a "one-stop" closed loop from multi-modal input to driving action output. Models can simultaneously recognize traffic signs, understand user commands, analyze complex road conditions, and output human-like coherent decisions, achieving the core user value of "low takeover rate and high trust".
SAIC IM's evolution path confirms the industry consensus: the implementation of L3 is not only the stacking of computing power and sensors but also the iteration of architectural logic-from "execution automation" of a single modality to "cognitive intelligence" of multi-modal fusion, and VLA is the ultimate carrier of this process.
From the iteration of consumer demand to the strategic positioning at the supply side, from the differentiation of technology paths to the upgrading of hardware computing power, the eight major development trends of L3 intelligent driving Are essentially a panoramic microcosm of the industry's leap from "intelligence 1.0" to "Intelligence 2.0". It is no longer a breakthrough in a single technology, but a systematic project of multi-dimensional collaboration of "demand - technology - commerce - ecosystem". In the future, L3 will not only be a "core weapon" for OEMs to break through homogenization and achieve brand upgrading but also a "key bridge" connecting L2 popularization and L4 ecosystem. When production vehicle models are intensively launched during 2025-2027, and the regulatory dividends of highway scenarios extend to urban areas, L3 intelligent driving will no longer be a synonym for "high-end configuration", but a "tipping point" that redefine automobile value and opens up a new smart mobility ecosystem.