Over the past decade, intelligence has primarily been realized in smartphones and PCs, with AI capabilities mostly residing in the cloud. However, the vast array of daily devices around ordinary people—lights and home appliances at home, desk lamps, earphones in our ears, and two-wheelers for commuting—although connected to the network, have never truly become "smart": network configuration is too complex, and users often fail to configure them properly, resulting in networked features that are rarely truly connected. Moreover, even when connected, they lack effective interaction and perception methods. As a result, while they can be remotely controlled via apps and report their status, they rarely understand what we say, what we are doing, or what we truly want.
Now, changes are beginning to occur. Large models have pushed the upper limit of cloud-based AI capabilities to unprecedented heights. Yet, the second half of AI is about bringing these capabilities truly down to the edge, into the consumer-grade terminals people interact with daily. This diffusion process has given rise to the concept of "Light Intelligence."
Light Intelligence aims to upgrade past "dumb" terminals into "four-haves" terminals: with connectivity, intelligence, perception, and interaction. Light Intelligence terminals can bear more understanding and judgment capabilities, realizing multimodal recognition, local inference, and real-time response, and collaborating with the cloud through connectivity to jointly complete an AI task.
From Two-Wheelers to AI Toys: Light Intelligence is Entering Real-World Application Scenarios
Whether "Light Intelligence" forms a trend depends on whether perception, connectivity, computing, and interaction capabilities truly enter mass-produced products, and whether manufacturers are willing to bear the added BOM and cloud costs in the long term.
Two-wheelers are a relatively typical case:
In the past, the electrification of two-wheelers mostly revolved around basic instruments, positioning, anti-theft, and smartphone apps. A typical approach to light intelligence retrofitting is to upgrade traditional segment LCD screens to smart dashboards, integrating navigation, networking, audio, and more interactive capabilities on a single screen. At the same time, NFC, 4G, and cameras are used to add unlocking, online services, and environmental perception. Cameras can also be used for riding records and safety warnings, while NearLink can support more precise digital car keys.
In the past, many smart functions of two-wheelers relied on smartphones to be completed, and the vehicles themselves remained merely execution terminals. When navigation, connectivity, display, and perception capabilities directly enter the vehicle end, the vehicle has the opportunity to become an independent mobile intelligent node. According to estimates, through a screen terminal retrofit costing around CNY 100, it is possible to build a more complete whole-vehicle intelligent experience.
For a large amount of mature hardware, intelligence does not necessarily require starting from scratch; it can also be upgraded by adding a layer of cost-controllable perception, connectivity, computing, and interaction capabilities. The key lies not in stacking functions, but in whether the added BOM can be translated into user-perceivable value in navigation, safety, interaction, or services.
Earphones, on the other hand, demonstrate another path:
The core tasks of traditional earphones are audio playback, calling, and noise cancellation. However, they naturally possess microphones and speakers and are worn close to the user for long periods, making them highly suitable as AI interaction entry points. Current upgrade directions include invoking cloud-based large models through edge-cloud collaboration to achieve semantic Q&A and real-time translation, as well as extracting human voice in noisy environments through local AI processing.
This kind of upgrade reflects the division of labor in "Light Intelligence": the edge handles high-frequency, real-time audio processing and interaction, while complex understanding is handed over to the cloud. Although earphones have limited computing power and battery life, they can leverage connectivity to gain greater AI capabilities, gradually transforming from pure audio peripherals into an entry point for user-AI interaction.
According to Counterpoint's forecast, the penetration rate of Edge AI in wearable devices will increase from about 30% in 2025 to nearly 80% in 2032, with TWS earphones and smartwatches remaining the two largest product categories in terms of shipment volume. The reasons behind this are not only the improvement in model capabilities but also the increased device integration, better performance per watt, and the growing user demand for real-time response and privacy protection.
Similar changes are quietly taking place in daily scenarios such as wearables, audio, smart homes, two-wheelers, AI toys, and AI desk lamps.
Why Are Light Intelligence Terminals Only Beginning to Scale Today?
First, the threshold for loading intelligence onto hardware is dropping. Connectivity, computing power, display, and perception capabilities are becoming increasingly accessible to ordinary terminals. OpenHarmony is helping the smart OS ecosystem foundation mature, AI coding is lowering the threshold for terminal development, and Agent + cloud models are making edge-cloud collaboration easier, achieving great intelligence with small computing power.
As both hardware and development costs decline simultaneously, the long-tail devices that were previously difficult to intelligentize due to low unit prices and fragmented categories now have a feasible path for intelligent implementation in new-generation product designs for the first time.
More importantly, changes come from the cloud. The development of large models and Agents poses new requirements for the edge: devices must possess perception, computing, and interaction capabilities matched to their own scenarios to truly bear Agents, acting as the hands and feet of the cloud "brain" to translate cloud intentions into real-world actions.
When hardware costs, edge computing, operating systems, and cloud AI simultaneously cross the usability threshold, those devices that "could only connect to the network" in the past are, for the first time, truly equipped with the conditions to upgrade to terminals that "can understand and respond."
From a Single Chip to the "Light Intelligence Foundation"
If the important hallmark of the last round of IoT upgrades was adding communication chips to traditional devices, then as AI enters billions of terminals, simply adding an NPU is obviously not enough.
For a device to truly possess intelligent capabilities, at least five issues must be resolved:
First is connectivity. In the past, IoT connectivity solved "whether a device can access the internet," and even now, some terminals remain unconnected. Light Intelligence must first solve "how devices collaborate with each other and how to connect to cloud AI." This includes both short-range communication and wide-area connections such as 4G and 5G. Taking NearLink as an example, HiSilicon has extended related applications to scenarios such as styluses, TVs, remote controls, earphones, digital car keys, and smart homes, providing SDKs, development boards, and reference cases.
Second are perception and interaction. Before AI enters the physical world, it must first be able to "see," "hear," and "express." Therefore, the technical foundation of AI terminals includes not only NPUs but also cameras, microphones, displays, audio, and various sensors.
This layer also includes audio and video interaction. The entry of ultra-high-definition audio and video into the Light Intelligence system is not merely about improving display specifications, but enabling terminals to "see," "hear," and "express" more naturally. GPMI has moved from a technical solution to the standardization stage. A series of industry standards for the "General Multimedia Interface Specification," including architecture, protocols, connectors and cables, and power supply, will be officially implemented on July 1, 2026.
The third layer is the operating system. One of the biggest differences between IoT and smartphones or PCs is the extremely fragmented device forms: a gas meter, a pair of earphones, a two-wheeler, and a TV face vastly different scenarios and need to process very different data. Therefore, the OS required for Light Intelligence must have a sufficiently large scalability range. As more and more terminals need to access AI, it also needs to shield hardware differences downwards, provide relatively unified application and AI interfaces upwards, and support multi-device collaboration. Otherwise, every time a new toy, home appliance, or wearable product is added, chip, driver, connectivity, and AI model adaptations must be redone, making it difficult for intelligence to truly scale.
The HarmonyOS ecosystem currently meets this requirement. According to data released by the OpenAtom Foundation in April 2026, the OpenHarmony ecosystem has over 1.3 billion devices, more than 500 partners, and over 1,700 products that have passed compatibility testing. Its flexible deployment capability can cover devices from the KB level to the GB level, extending from sensors and controllers to complex interactive terminals.
The fourth layer is edge AI. Many past IoT terminals might not have possessed intelligent capabilities. Currently, Edge AI processors can support models such as CNN, RNN, and Transformer; lower-resource embedded terminals can use lighter AI acceleration capabilities. People are paying more attention to how much truly effective AI capability each watt of power consumption and each CNY of BOM can provide.
The final link is development cost. For a company making desk lamps, toys, or small home appliances, the real obstacle may not be the inability to afford AI chips, but whether they have the capability to complete model adaptation, driver, connectivity, cloud interface, and application development. In the global edge AI competition, the importance of SDKs, model conversion, IDEs, and reference designs is rising. Taking the HiSilicon Developer Ecosystem Community as an example, it has publicly provided SDKs such as WS63, BS2X, and Hi3403, as well as HiSpark Studio, HiSpark AI, burning tools, and compilation toolchains.
This also reflects the industry trend of chip companies moving from "selling chips" to "providing foundations." HiSilicon's layout in Light Intelligence is a sample of this trend: combining the operating system, connectivity, edge AI, audio and video capabilities, and development tools into a low-threshold infrastructure, allowing devices with different cost, power, and functional positioning to find corresponding upgrade paths. Consequently, the unit of competition expands from a single SoC to a combination of chips, software, toolchains, and ecosystems.
Conclusion: Large Models Determine the Upper Limit, Light Intelligence Determines How Far AI Can Go
Finally, over the past few years, the AI industry has continuously scaled up models and pushed computing power higher. However, as AI begins to enter the more massive number of IoT terminals, the problems faced by the industry are also changing: how to enable more devices to gain truly valuable intelligence at lower costs, lower power consumption, and lower development thresholds.
If the first half of AI solved the "upper limit of capabilities," then the more important issue in the next stage is to continuously lower the "entry threshold" for intelligent capabilities. Large models determine how smart AI can be, while Light Intelligence determines how many devices and scenarios AI can enter.
From an industry perspective, HiSilicon's layout around OpenHarmony, NearLink connectivity, edge AI, ultra-high-definition audio and video, and the HiSpark developer community represents a path for chip manufacturers to extend from single-point products to platform-based capabilities. For terminal manufacturers, what is truly valuable is not the addition of technical terms, but whether these capabilities can be invoked at a lower cost and further diffused into consumer electronics, automotive electronics, and industrial equipment.
It can be seen that "Light Intelligence" is becoming an industry consensus. The competition in edge AI in the next stage is not about who has the larger chip computing power, but who can enable customers to achieve intelligence at a lower cost, higher efficiency, and with the highest security.
Editor: Langkejian, XinzhiXun