ZoZo Auto Research has released the "2026 Automotive AIOS Research Report".
Small-Scale Deployment of Mass Production Solutions
In 2026, AIOS begins to see small-scale deployment, helping to refine various cabin AI functions and expand application scenarios. Furthermore, the AIOS in some mainstream flagship models implements cross-domain orchestration capabilities through atomic services, extending execution capabilities to domains such as body, chassis, and autonomous driving.
OEM Models Featuring Mass-Produced AIOS
Source: ZoZo Auto Research
As of June 2026, OEMs' AIOS R&D approaches are still divided into two methods: in-house development and semi-outsourcing:
Full-stack in-house development: Emerging carmakers led by NIO and Li Auto deeply integrate AI capabilities into the middleware layer (or even the kernel layer), achieving a full-stack closed loop from chips to applications.
Semi-outsourcing: Traditional OEMs and other manufacturers independently build domain-specific large models + AIOS frameworks, reusing suppliers' basic software at the underlying layer.
Supplier Solutions for Mass-Produced AIOS
Compiled by: ZoZo Auto Research
The integration methods for suppliers' mass-produced AIOS solutions include the following:
Extending from cabin AI applications to the underlying OS: The mainstream approach, such as Huawei providing HarmonyOS-related services through the Harmony Cockpit.
Binding with chip/hardware manufacturers: Adapting to chip solutions from multiple chip manufacturers to build hardware-software synergy. Typical examples include SenseTime's Sage Box for SageOS and ThunderSoft's AI Box-N1 for DiShui OS. By collaborating with AI Box, they build comprehensive edge-side AI solutions.
Binding with cloud providers: Represented by Yitu Technology, binding with Volcengine's cloud infrastructure, invoking the Doubao large model to provide AI services.
Compared to 2025, OS cross-domain invocation services are more refined in 2026.
Taking Dongfeng Tianyuan OS as an example, the entire architecture achieves a five-in-one cross-domain integration of body, powertrain, chassis, thermal management, and gateway. Based on the "Taiji Large Model" foundation, it invokes over 2,000 atomic services, supporting rapid combination and flexible invocation of functions through a service-oriented architecture.
Dongfeng Tianyuan OS Architecture
Source: Dongfeng Motor
During the deployment of AIOS, the software foundation for cross-domain invocation in the cabin software system remains the vehicle-level OS. Based on the vehicle-level OS, atomic decomposition of non-safety-related cabin functions can be completed. Then, through automotive-grade intelligent scheduling algorithms deployed in the AI middleware, dynamic allocation and intelligent scheduling of computing power, applications, and peripheral resources within the cabin are realized. When a user issues a command, the voice assistant decomposes the intent, coordinates the work of multiple AI frameworks, and ultimately invokes atomic services. This entire process constitutes the basic workflow of the AIOS system.
In 2026, the number of interfaces for atomic capabilities has been greatly expanded (the number of atomic capabilities in mainstream flagship models basically exceeds 500+). Against the backdrop of increasingly prevalent customized cabin scenarios, the competitiveness of AIOS is gradually shifting from "having more atomic capabilities" to "whether atomic capabilities are easier to combine." Regarding "whether they are easier to combine," customized interfaces and standardized interface protocols demonstrate their importance.
Some engineering challenges that highlight the effects of AIOS are as follows:
AIOS Engineering Challenges (Partial)
Compiled by: ZoZo Auto Research
Among these, the lack of a standardized unified protocol can easily drag down the effectiveness of atomic capabilities in the SOA architecture. Currently, the relevant protocols still mainly adopt Function Call, while MCP is in a trial stage. The reason is that in the context of small-scale mass production, Function Call can still meet the main needs and is easy to maintain; however, when solutions need to be migrated on a large scale, wrapping a layer of MCP Server outside the Function Call demonstrates the advantages of "cross-model portability" and "dynamic tool discovery."
The significance of the MCP protocol lies in standardization, compressing the cross-scenario function development cycle from months to weeks. Automakers can quickly combine "highly personalized" cabin services like building blocks. Typical representative cases include Yitu Technology's in-vehicle MCP-Agent framework and SenseTime Jueying's edge-native agent framework supporting the MCP/A2A protocol.
Taking SenseTime Jueying as an example, it has launched an edge-native agent framework supporting the MCP/A2A protocol, building a standardized "agent-tool" docking framework. This allows multiple agents to efficiently dock with various in-vehicle tools such as players, air conditioners, and knowledge bases through a unified MCP protocol layer, solving the difficulties of complex tool invocation, difficult data acquisition, and multi-source information integration during agent development.
Its advantages include:
Reducing costs and improving efficiency. The unified protocol eliminates the fragmentation barriers in tool docking, significantly reducing development and collaboration costs, enabling various tools to be "plug and play."
Open ecosystem. It supports standardized ecosystem access mechanisms, facilitating the rapid integration of third-party services and hardware into the in-vehicle intelligent system, and promoting ecosystem diversification.
Secure and controllable. It unifies security authentication strategies and manages them centrally, simplifying processes while strengthening system security.
SenseTime Jueying's Edge-Native Agent Framework for MCP/A2A Protocols
Source: SenseTime Jueying
Next Phase: Shifting from the Driven Stage to the Native Stage
Suppliers such as Huawei and Neusoft propose that the integration of AI and OS is divided into three stages:
Three Stages of AIOS Evolution
Source: Public Channels
AIOS Evolution Characteristics Proposed by Huawei
Source: Huawei
In 2025, most OEMs and suppliers adopted the approach of deploying AI frameworks in the middleware layer to build AI operating systems; for example, XPeng deployed a cross-domain unified protocol middleware + edge-side large model + atomic services, Great Wall Motor deployed a multi-model foundation and Agent management/operation framework in the middleware layer, and Neusoft Reach's NeuSAR OS deployed the NeuSAR AI Framework for the rapid integration of AI applications.
In 2026, leading emerging OEMs and suppliers began to deploy "AI-enhanced kernels," attempting to build native AIOS at the kernel layer; for example, NIO uses AI to enhance the OS kernel's ability to dynamically schedule resources based on scenarios, and the kernel layer of Huawei's HarmonyOS natively supports multimodal understanding and personalized data understanding.
AIOS Shifting from the Driven Stage to the Native Stage
Source: ZoZo Auto Research
Furthermore, with the deployment of Agent technology and the upgrade of high-computing-power chips, the AIOS architecture has also changed from the application layer to the underlying layer:
Changes in AIOS Architecture from Application Layer to Underlying Layer
Source: ZoZo Auto Research
Taking NIO's new SkyOS as an example, the system deeply integrates AI capabilities into the underlying layer of the operating system, replacing the traditional architecture paradigm. It achieves highly efficient edge-cloud collaboration, intelligent scheduling of heterogeneous computing power, and multi-agent collaboration, while improving the system's response speed, stability, data throughput, and battery endurance. The innovation lies in enhancing the OS kernel's ability to dynamically schedule resources based on scenarios, including CPU (process and thread priority) and memory management (allocation and reclamation), device sharing, etc., which can improve system stability and response speed under high-load scenarios.
New SkyOS Architecture Supports AI-Enhanced Task Scheduling and Other Functions
Source: NIO
Taking Huawei Qiankun OS as an example, it includes a security isolation engine, AI-native kernel, Lingqu bus, acceleration engine, and BiSheng compiler, supporting deterministic low-latency fast response for upper-layer ADS algorithms. The kernel is based on the HarmonyOS kernel but has been deeply tailored and reconstructed for automotive scenarios, becoming an AI-native kernel that can also achieve seamless flow between "vehicle-road-cloud-mobile phone."
Huawei Qiankun OS
Source: Huawei
[Contact Information] Mobile number is the same as WeChat number
Industry Research Department | Mr. Fu 15810027571
Mr. Zhao 18702148304
Data Service Department | Ms. Zhang 13716037793
Strategic Consulting Department | Ms. Han 15810133447
Promotion and Communication Department | Ms. Liao 13718845418
Mr. Du 13910162318