CAICT, together with Huawei, vivo, StepFun and others, releases the "Device-Side Intelligent Industry Development Report (2026)"

On October 14, 2026, at the "On-Device Intelligence Forum" of the Artificial Intelligence Industry Conference, the Artificial Intelligence Research Institute of the China Academy of Information and Communications Technology (CAICT), together with the On-Device Intelligence Group of the Artificial Intelligence Industry Development Alliance of China, officially released the "On-Device Intelligence Industry Development Report (2026)".

The China Academy of Information and Communications Technology, together with Huawei, ZTE, vivo, Haier, StepFun, ModelBest, Dahua Technology and other organizations, jointly released the "On-Device Intelligence Industry Development Report (2026)".

Yan Shu, Deputy Chief Engineer of the Artificial Intelligence Research Institute of CAICT, presented an interpretation of the report.

At present, artificial intelligence is accelerating its evolution from centralized cloud computing toward device-cloud collaboration, ubiquitous deployment and scenario-based services. As large models, multimodal models and agent technologies continue to make breakthroughs, model capabilities are constantly extending toward the data generation side, the user interaction side and the task execution side, and on-device intelligence has become an important direction for the deep integration of artificial intelligence with the physical world. At the same time, technologies such as model lightweighting, heterogeneous computing, efficient storage, inference optimization and foundational software continue to develop, enabling artificial intelligence to be deployed efficiently across more diverse hardware resources and operating environments.

The report systematically reviews the development trajectory, policy environment, market dynamics and conceptual characteristics of the on-device intelligence industry, with a focus on key technologies such as model lightweighting, heterogeneous hardware integration, foundational software, agents, device-cloud collaboration and security and trustworthiness. It examines native product forms such as personal intelligence, spatial intelligence, mobile intelligence, industry intelligence nodes and embodied intelligence, reviews the development characteristics of consumer entertainment, smart home, smart office, smart travel, smart health and integrated scenarios, and further assesses development trends, main challenges and advancement paths for on-device intelligence in terms of technology, applications, ecosystem and security, providing a reference for technological innovation, industrial layout, standards development and application promotion. Core viewpoints of the report

1. On-device intelligence is a fused intelligence form that is primarily local and involves device-cloud collaboration. On-device intelligence takes terminal devices as its entry point, prioritizes the use of local software and hardware resources to complete localized intelligent services of perception, inference, decision-making and execution, and combines cloud computing power, models and knowledge capabilities to complete complex tasks in a fused intelligence form. In a broad sense, on-device intelligence is not equivalent to simply "running models locally"; rather, it takes local intelligence as its capability foundation and device-cloud collaboration as its operating mode, covering a full-stack capability system encompassing models, chips, sensors, storage, foundational software, agent frameworks, application services, security mechanisms and operational ecosystems.

2. Model lightweighting and software-hardware collaboration have become the main line of technological evolution for on-device intelligence, with technological competition shifting from single-point performance improvement to optimization of overall system operating efficiency. Faced with resource constraints such as computing power, storage, bandwidth and power consumption, on-device models are improving capability density under limited resources through efficient architectures, quantization, distillation, sparsification, inference optimization and collaboration between large and small models; heterogeneous computing such as CPU, GPU and NPU, efficient storage and low-power perception provide the foundation for model operation; compilers, runtimes and foundational software further undertake model conversion, operator mapping, resource scheduling, task orchestration and device-cloud collaboration. On-device intelligence is moving from "being able to deploy" to "operating efficiently, stably and continuously", and the focus of technological competition is shifting from single-model scale or peak computing power to the system collaboration efficiency of models, hardware, foundational software and operating environments.

3. The deep integration of model capabilities with diverse hardware is giving rise to intelligence-native products that differ from traditional "hardware + functions". Intelligence-native products organize software and hardware resources around perception, model, memory, inference and execution capabilities from the design stage; models are no longer merely additional functions on existing products, but are gradually participating in interaction entry points, system scheduling and task loops. Depending on differences in usage environments, real-time requirements and execution methods, forms such as personal intelligence, spatial intelligence, mobile intelligence, industry intelligence nodes and embodied intelligence are developing rapidly. The differences between product forms essentially stem from different combinations of model capabilities with computing power, storage, sensing, interaction and execution resources, and product innovation is shifting from function stacking to collaborative design of models and hardware.

4. On-device intelligence is accelerating its penetration into multiple types of real-world scenarios, with application methods expanding from single-point functions to continuous services, multi-device collaboration and task loops. On-device intelligence is accelerating its penetration into scenarios such as consumer entertainment, smart home, smart office, smart travel and smart health, driving artificial intelligence capabilities from single-point functional applications toward deep embedding in users' daily life and production activities. Different scenarios impose differentiated requirements on on-device intelligence capabilities: consumer entertainment scenarios focus more on content generation, real-time interaction and personalized experiences; smart home emphasizes multi-device linkage, environmental perception and proactive services; smart office focuses on knowledge processing, task collaboration and workflow reshaping; smart travel highlights low-latency perception, continuous interaction and safety assistance; and smart health relies on continuous perception, privacy protection and long-term health management. As on-device computing power, lightweight models, multimodal perception and agent capabilities continue to improve, on-device intelligence applications are evolving from "single device, single function, single scenario" to "multiple devices, multiple tasks, cross-scenario", and service methods are gradually shifting from passive response to proactive perception, proactive recommendation and proactive execution.

5. On-device intelligence is entering a new stage, moving from "whether a model can be deployed" to "whether models, hardware, systems and services can work together." Lightweight models continue to increase capability density per unit of resources; heterogeneous hardware and foundational software translate compression gains into real operating efficiency; and AI-native products further embed local perception, inference and execution into real tasks. At the same time, continuous intelligent services, cross-device collaboration and agent-based business forms are taking shape at an accelerating pace, and industrial value is expanding from incremental hardware functionality toward incremental capabilities, services and ecosystems.

6. Technology and products are evolving in tandem, and on-device intelligence will accelerate from single-point capability breakthroughs toward large-scale applications and full-stack ecosystem collaboration. At the technology level, model lightweighting and software-hardware collaboration will continue to deepen, and on-device intelligence will evolve from "being deployable" to "operating efficiently, stably and continuously"; at the application level, artificial intelligence will move further from feature enhancement toward AI-native operation and closed-loop tasks, interaction methods will gradually shift from "operating functions" to "expressing goals", and continuous services, cross-device collaboration and agents' autonomous execution capabilities will continue to strengthen; at the ecosystem level, industrial competition will shift from single-point product capabilities to full-stack collaboration across models, hardware, foundational software, products and services, new links such as model deployment, development tools, capability platforms, agent services and operational services will continue to expand, and industrial value will extend from one-time product sales toward continuous services and ecosystem operations. At the same time, security governance will gradually be integrated into the full lifecycle of products and services alongside the development of models, agents and actual execution capabilities, providing a foundational safeguard for the large-scale application of on-device intelligence. Report Table of Contents

I. Overview of the On-Device Intelligence Industry

(1) The industry is moving from feature embedding toward system intelligence, and agents are driving accelerated leaps in on-device capabilities

(2) Policies continue to strengthen the penetration of artificial intelligence into on-device and physical scenarios

(3) Industrial value is expanding from incremental hardware toward incremental intelligent capabilities and services

(4) On-device intelligence is a fused intelligence form that is primarily local and features device-cloud collaboration

II. On-Device Intelligence Technology System

(1) Intelligent hardware layer: heterogeneous computing and efficient storage support stable model operation

(2) Foundational software layer: compilation adaptation and resource scheduling open up the execution path from models to hardware

(3) On-device model layer: lightweighting and efficient inference improve capability density under limited resources

(5) On-device security and trust layer: building a trusted boundary covering models, data and agent execution

III. AI-Native Product Forms for On-Device Intelligence

(1) Personal intelligent products: collaboration among multiple types of hardware forms intelligent entry points

(2) Spatial intelligent products: the integration of perception, decision-making and execution forms environment-level intelligence

(3) Mobile intelligent products: local real-time inference supports autonomous perception and mobile decision-making

(4) Industry intelligent nodes: the downward deployment of lightweight models drives field devices from collection toward decision-making

(5) Embodied intelligent products: the integration of models with sensing, control and execution systems moves toward closed-loop physical tasks

IV. Application Scenarios for On-Device Intelligence

(1) Consumer entertainment: centered on content services, with ultimate user experience driving commercial conversion

(2) Smart home: centered on multi-device collaboration, moving from passive response toward a new model of proactive service

(3) Smart office: centered on improving efficiency, reshaping knowledge management and workflow systems

(4) Smart travel: centered on safety services, with travel efficiency and interaction experience continuously optimized

(5) Smart health: centered on continuous perception, building a health service system integrating monitoring, detection and intervention

(6) Integrated scenarios: connecting multi-device service links to build all-scenario intelligent collaboration

V. Summary and Outlook

(1) Trends: technology and products evolve in tandem, and on-device intelligence accelerates toward large-scale implementation

(2) Challenges: multiple bottlenecks intertwine and overlap, and industry maturity still requires systemic breakthroughs

(3) Recommendations: strengthen technology-ecosystem collaboration and consolidate the foundation for on-device intelligence development. Brief Introduction of Key Experts

Jiang Chunyu, Senior Engineer and Director of the Big Data and Intelligence Department of the Institute of Artificial Intelligence at the China Academy of Information and Communications Technology (CAICT)

Chair of the Big Data Technology Standards Promotion Committee of the China Communications Standards Association. His research directions include big data, artificial intelligence technology, large models, data element circulation, data asset management and data security. He participated in drafting multiple national documents, including the Action Outline for Promoting National Big Data Development and the Big Data Industry Development Plan (2016-2020), undertook major research projects and industrial development support work for multiple government departments, and led the development of multiple standards and specifications and the compilation of industrial research reports.

Pang Weiwei, Engineer and Deputy Director of the Big Data and Intelligence Department of the Institute of Artificial Intelligence at CAICT

He has long been engaged in research on technologies and applications such as on-device intelligence, digital identity and blockchain. He led or participated in multiple project topics under the National Key R&D Program and has produced multiple achievements, including industry research reports, standards and specifications, papers and patents.

Liu Tingting, Engineer at the Institute of Artificial Intelligence at CAICT

She is mainly engaged in technical research and standardization work in fields such as on-device intelligence, digital identity and digital trust. She led the development of multiple international standards, including ITU-T and IEEE standards, participated in multiple National Key R&D Program projects, and has produced multiple achievements, including industry research reports, papers and patents.

Zhang Lifeng, Engineer at the Institute of Artificial Intelligence at CAICT

Secretary-General of the IEEE Digital Trust Standards Committee and Secretary of the Trust Technology Global Promotion Plan. He is mainly engaged in technical research and standardization work in fields such as on-device intelligence and digital identity. He led the development of more than 10 international standards, including ITU-T and IEEE standards, participated in multiple National Key R&D Program projects, and has published multiple SCI- and EI-indexed journal papers as corresponding author.

Source: China Academy of Information and Communications Technology and Artificial Intelligence Industry Alliance of China Report Download

Follow the official WeChat account of this site, "Smart City Industry Analysis," and enter the keyword "CAICT AI Report" in the chat box to obtain the standard download address: www.smartcity.team

Statement: Please indicate the source as "Smart City Industry Analysis (www.smartcity.team)" and the relevant author when reprinting. Except for original content, some information on this site comes from the Internet. If the relevant content involves infringement of rights, please contact us via the contact information at the bottom to delete it.

[Digital Technology Special Research] Community

In-depth peer exchange, focused on research and material accumulation in the smart city and digital technology industries.

Continuously updated with the latest industry white papers, industry research, policy developments, and implementation solutions, helping industry practitioners efficiently accumulate knowledge and share experience.

Massive amounts of curated content are continuously iterated, providing cutting-edge perspectives, connecting industry talent, carrying out collaborative case creation and resource exchange, and jointly building a high-quality industry exchange atmosphere.

#post-##

Recommended for you

"Green Token Factory Technical White Paper": Systematically elaborates the five-layer technical system of the green token factory: "energy infrastructure—computing infrastructure—token production—token operations—token circulation"

Guidelines for Building High-Quality Datasets for Agent Capability Construction (2026): Jointly released by CAICT, Lenovo, and the AIIA Data Committee, proposing the overall framework, key methods, and implementation paths for building industry high-quality datasets for agent capability construction

"Research Report on the Development of Intelligent Computing Infrastructure for Token Services (2026)" by the Planning Institute of CAICT: Systematically reviews the paradigm reconstruction process of intelligent computing infrastructure from "computing factories" to "token factories"

Research Report on the Development of the Token Economy (2026), with an accompanying PPT interpretation by Li Wei, Director of the Cloud Computing and Big Data Institute

"Engineering of Stewardship: Research Report on Empowering Intelligent-Native Software Engineering (2026)" by the Cloud Computing and Big Data Institute of CAICT

China Embodied Intelligence Data Industry Analysis Report 2026: Detailed explanation of the current status, ecosystem, business models, typical enterprises, and future development trends of China's embodied intelligence data industry