Author | Chen Junda Editor | Mo Ying
Over the past few years, AI development has been concentrated in the cloud: more parameters, more powerful models, and increasingly power-hungry computing centers. Today, a deeper shift is underway: AI is beginning to visibly diffuse from the cloud to edge devices. An increasing number of terminals, such as robots, autonomous vehicles, and smart vision devices, are acquiring their own AI capabilities.
This trend is not hard to understand. Terminals are the critical gateway for AI to reach the real world: cameras, microphones, and sensors are built directly into the devices, continuously generating data from the physical world. No matter how powerful the cloud models and computing power are, if they cannot reach the terminals, they cannot touch the real world. It can be said that the next battlefield for AI competition is right at the edge.
At this juncture, Shanghai HiSilicon has made a clear judgment: every IoT terminal deserves to be reconstructed into a light intelligence terminal, becoming a ubiquitous gateway for AI. Through the collaboration of edge AI and cloud, it achieves great intelligence with small computing power. Centered on this judgment, a complete architecture from strategy to execution has been rolled out: on the one hand, leveraging a full series of edge chips to capture the opportunity of the edge AI industry explosion; on the other hand, accelerating the construction of a new-generation developer ecosystem for edge AI through the HiSpark developer ecosystem community.
01.The Edge Has Entered the Era of Light Intelligence
A common consensus is that as cloud large models become increasingly powerful, the next stage of AI development will inevitably move towards the real physical world. In this process, several hard constraints will certainly be encountered.
First is real-time response. Taking a typical cloud command as an example, the round-trip communication time for a common cloud voice command takes at least about 1000 milliseconds. This is fine for normal interactions, but high-speed decision-making scenarios with extreme latency pressure simply cannot tolerate such delays.
Second is multimodal perception: cameras, microphones, and sensors are built directly into the devices, enabling terminals to continuously and real-timely acquire data from the physical world, which the cloud cannot obtain. Finally, considering privacy protection and data security, keeping data and commands within the terminal naturally avoids a large number of risks. Therefore, deep collaboration between the cloud side and the terminal side is required to build a complete business chain.
The stronger the edge capabilities, the more data can be collected and processed, and the higher the demand for the cloud. The improvement of cloud model capabilities will in turn further drive terminal upgrades, forming a continuously enhancing synergistic relationship. Shanghai HiSilicon defines this terminal evolution trend as "Light Intelligence." The so-called "light" means enabling terminal devices to achieve "connectivity, intelligence, perception, and interaction" in one stop with lower cost and power consumption, changing the past situation where AIoT terminals lacked true "AI." In particular, the introduction of edge AI capabilities redefines terminal products, making the brand-new "light intelligence" terminals the largest carrier for AI to truly enter daily life, and also a key to completely activating the AIoT industry.
This change will ultimately fall on specific product experiences. Let us try to imagine equipping a smart toothbrush with a micro camera and an edge computing chip, enabling it to scan teeth while brushing, determining where focused cleaning is needed or even if medical attention is required; a hair dryer is no longer a functional product with fixed hot and cold gears, but a complete smart product that can dynamically adjust wind temperature and speed according to hair quality and wetness. Shan Gang, Director of the Ecosystem and Partner Development Department of Shanghai HiSilicon, believes that a light intelligence terminal is not about creating a brand-new "species," but about endowing traditional products that consumers are already familiar with with brand-new, generation-defining intelligent capabilities.
02.Computing Power is Easily Accessible, but Implementation is Difficult Ecosystem Support Becomes a Practical Need
However, whether edge AI can run depends not only on chip computing power. Shan Gang pointed out that edge AI faces a very practical difficulty: "Computing power is easily accessible, but implementation is difficult." Integrating computing power is easy and hardware costs are low, but after enterprises obtain the computing power, the real test has just begun.
The first hurdle is "not being able to figure out" what the computing power can do. This requires enterprises to have profound industry know-how, sufficient insight and cognition of the functional scenarios and deep-level user needs of products in their industry, to deeply understand the problem, and to reasonably define the experience. This hurdle alone keeps a large number of enterprises out. And even if they figure out what the computing power can do, the second hurdle is to figure out how to implement the model. Edge AI involves a whole set of interlocking engineering chains. For a large number of small and medium-sized enterprises, what is truly scarce is not computing power, but the ability to turn computing power into products.
In embedded development, developers always have to find a balance between computing power, power consumption, and cost: if computing power is high, power consumption and heat generation may rise; if power consumption is reduced, performance may be sacrificed. There are strict constraints on memory, power consumption, computing power, and cost. On the software and model side, how to finally deploy cloud-trained models to edge devices through quantization, pruning, compression, and other methods, and how to optimize inference frameworks and underlying operators for different chips, are extremely complex challenges. This set of software and hardware adaptation and algorithm R&D capabilities was more in the hands of large companies with large-scale R&D teams in the past.
However, edge AI is a naturally fragmented market, and many participants are small and medium-sized enterprises, solution providers, and individual developers. For them, the cost of independently going through the entire chain is unbearable. So, where is the way forward for the vast number of developers? The rise of AI programming seems to be providing new path inspiration for lowering the threshold of edge AI development engineering. Logically, developers can let large models generate code, call tools, and even assist in completing the entire development process. But embedded development faces specific chips, specific hardware, and specific engineering environments, and general large models cannot directly solve these problems. They may not know the BSP, SDK, and signal pins of specific chips, and it is difficult for them to keep up with constantly updated versions, interfaces, and errata.
More importantly, embedded development is not just about writing code. The model must also understand hardware constraints such as computing power, power consumption, and cost. Even if the generated code is logically correct, it may not be able to truly run on the board due to memory, performance, and other issues; during the debugging stage, tools such as logs, logic analyzers, and oscilloscopes must be connected to form a complete "human-model-hardware" closed loop. This means that general AI programming can lower some development thresholds, but it cannot independently make up for the engineering system required for edge AI implementation.
How can most enterprises obtain the systems engineering capabilities to build edge AI products? This is a question that must be answered in the wave of light intelligence, and it is also the proposition Shanghai HiSilicon has set for itself—after all, no matter how powerful the chip is, if massive developers cannot use it, it is in vain. To answer this proposition, relying on the chip itself is not enough. Since the bottleneck is in the engineering chain, the main battlefield of competition will inevitably extend from a single chip or model capability to the toolchain and developer ecosystem, and the role of chip manufacturers must change accordingly. In the past, chip manufacturers provided chips, SDKs, and technical support; today, developers need full-chain support from chips, models, and toolchains to demos, and manufacturers need to accompany developers through the entire process from computing power to products, ultimately becoming providers of edge AI development platforms.
This is also the direction Shanghai HiSilicon is promoting: with the four major ecosystems as the foundation, providing rich chip products and solutions, continuously opening up chips, toolchains, models, and engineering capabilities, doing a good job in building the silicon-based fertile ground, so that more partners and developers can release their innovation potential on this fertile land. This round of upgrades to the HiSpark community is also carried out around this idea, extending from providing chips and SDKs further towards a complete edge AI development ecosystem.
03.Pioneering the XDAS Four-Dimensional Framework Remaking the Developer Ecosystem for AI
Shanghai HiSilicon may be one of the original manufacturers with the highest completeness and richness of chip sequences accessible to domestic developers. This trump card is the full-series, full-tier chip family: chips for perception, connectivity, display, and computing power are gradually deepening open-source and openness to developers, and different chips are connected through a unified operator library and a unified toolchain, allowing developers to migrate between different computing power tiers at low cost. This unification also extends to the edge-cloud: Shanghai HiSilicon's edge platform shares the same architecture with the CANN system. CANN is the full-stack development ecosystem of the Ascend system, and the edge chips are equivalent to extending this ecosystem all the way to the smallest smart terminals.
As the ecosystem work continues to advance, from the recently released Ascend 960 to HiSilicon's Hi3516CV610 with only 1T computing power, all are expected to share the same operator library, tool platform, model library, and inference framework in the future. On top of the chip foundation, the HiSpark community has proposed the XDAS developer ecosystem four-dimensional framework for the first time, systematically summarizing ecosystem capabilities into four dimensions: AI for Development (D), AI for Application (A), AI for Support (S), and EXperience for AI (X). The first three dimensions solve "how developers use AI," and X solves "how AI uses the community"—when the service objects of the community expand from humans to AI, the infrastructure of the ecosystem itself must also be remade.
Breaking it down, in terms of AI for Development, Shanghai HiSilicon is helping developers use AI to improve embedded development efficiency. They have structured and vectorized all SDKs, interface documents, reference projects, and successful experiences to build a real-time refreshing exclusive knowledge base, and then abstracted chip development capabilities into an atomic set of Skills released through MCP services and CLI tool links to access real development links. With this knowledge, processes such as code generation, compilation, burning, to reading logs and instrument data can all be handed over to AI for automatic iteration, with humans only responsible for raising requirements at the beginning and acceptance at the end. The first batch of Skills in five major domains created by Shanghai HiSilicon has been fully opened to the HiSpark community. Laboratory tests show an efficiency improvement of about 76%. Workloads that previously required one engineer about two weeks can now be completed in one day with the help of AI Skills.
Facing AI for Application, that is, application implementation, Shanghai HiSilicon innovatively proposed the "five-step method" problem-solving idea to help developers achieve the final step from idea to product. The HiSpark community breaks down application implementation into five links: "Case → Algorithm → Data → Training → Deployment", providing targeted support. The first step of the journey starts from the case center, where developers can learn existing cases, read development guides, and get development inspiration. With ideas, they can go to HiSpark's model and algorithm market to match needs. It provides a large number of open-source algorithms and models. Currently, the Hi3403 chip has opened 15 categories and over 70 open-source algorithms, and the Hi3516CV610, with annual shipments of over 100 million, will also open over 40 algorithms in the next three months.
Next, entering the training and deployment links, the HiSpark community provides tools such as model quantization pruning, compression, and conversion. By the end of this year, an online training platform will also be launched. Small teams no longer need to build their own training servers; they can get the model by uploading data. The five-step method paves the way from idea to product, but when developers actually get their hands dirty, they will still get stuck on countless specific problems.
How to make technical support faster and more accurate? This is exactly the problem that "AI for Support (S)" in the XDAS framework aims to solve. As the edge AI development ecosystem continues to enrich, traditional human FAEs have to deal with more and more problems and often cannot respond at the fastest speed. Internal statistics from Shanghai HiSilicon found that over 85% of the problems encountered by developers are actually repetitive problems that have been solved by predecessors. Therefore, the HiSpark community recently officially launched the digital FAE beta test, providing 7x24-hour services covering scenarios such as chip selection, technical guidance, data retrieval, root cause analysis, and solutions. It can also access developers' own IDEs through MCP; when encountering problems that AI cannot solve, the system will automatically organize the dialogue and transfer it to a human FAE.
Shanghai HiSilicon specifically emphasizes that the digital FAE is not an ordinary intelligent customer service: developers value accuracy and efficiency more, and an incorrect hardware capability judgment may bring high trial-and-error costs. Therefore, the system would rather not answer than answer randomly, relying on a strictly defined knowledge system and continuous errata behind it. And as developers start to use AI more and more, a new problem emerges: Is the community itself AI-friendly? This corresponds to the last dimension of the XDAS framework—EXperience for AI (X). HiSpark community data shows that the application volume for Markdown format documents has exceeded half, of which nearly 70% comes from agents. The service objects of the community are gradually shifting from "humans" to "AI."
To this end, the HiSpark community, on the one hand, requires documents, SDKs, interfaces, and engineering references to be "readable by humans and also readable by AI," and on the other hand, opens interfaces so that developers can directly call community capabilities in their own Agents. At the same time, it ensures information is accurate, clean, and updated in real time through data governance, avoiding the repeated dissemination of incorrect knowledge by AI. The reconstruction of the above four dimensions is bringing visible changes. As of September 2026, the download volume of the HiSpark community has increased by 10 times compared to last year, and UV has increased by 8 times. However, Shanghai HiSilicon's construction goal for the community does not focus on the number of developers or DAU, but values community health, product experience, and whether partners can truly succeed. Shan Gang emphasized that the value of the community is not "acquiring developers," but "serving and supporting developers." In the future, the community does not even need to exist as an independent website. Digital FAE, MCP services, and development capabilities can directly enter development environments such as IDEs and Agents. This is perhaps the state where the edge AI ecosystem is truly integrated into the development process.
04.Conclusion: Building a Path for Edge AI That Everyone Can Walk
The era of light intelligence has arrived, and this is forming a resonance within the industry. From traditional security to smart homes, from consumer electronics to industrial manufacturing, a large number of practitioners have re-seen opportunities in this wave. China has the richest terminal categories and the most active development force in the world, and tens of thousands of small and medium-sized enterprises are the most extensive participants in this market.
In the past, intelligent upgrades often meant high thresholds and high investments, which only head enterprises could afford; but when the ecosystem lowers the threshold layer by layer, allowing a small team to also make AI products, the dividends of technology truly begin to flow to thousands of industries and benefit thousands of factories and merchants. This is exactly what Shanghai HiSilicon is doing. With a full series of chips as the foundation and the HiSpark developer ecosystem community as the carrier, it comprehensively opens up knowledge, tools, models, and services, attempting to build a path that everyone can walk for China's edge AI industry.