Desktop companion robots, wall-mounted interactive smart paintings, wrist-worn voice agents... Today, the boundaries of AI hardware are constantly expanding driven by developers' creativity, and devices are gradually evolving from "connected terminals" into "intelligent agents capable of perceiving, understanding, and acting."
Meanwhile, a core issue is becoming increasingly prominent: developing an AI hardware product often requires a team to tackle the formidable challenges of chip drivers, inference frameworks, model compression, power consumption optimization, and OTA upgrades simultaneously. An "intelligent foundation" that can consolidate common capabilities into open infrastructure is becoming more and more necessary.
Recently, the Xiaomi openvela AI Hardware Innovation Summit, a supporting enterprise ecosystem event for the first Immersive Consumer Carnival, was held at the Beijing Tongming Lake Convention and Exhibition Center. Attendees engaged in discussions on core topics such as the openvela system architecture—the "intelligent foundation" for hardware in the AI era—industrial applications, and developer practices, providing a window to observe the path of AI hardware from "concept to product, and from functional to user-friendly."
Consolidating Common Capabilities to Enable Developers to Focus on Innovation
As ideas for hardware development continue to emerge, the forms, categories, and functions of current AI terminals are becoming increasingly diverse. At the event, Xu Shengqi, founder of Shanghai Starlight Gravity Digital Technology Co., Ltd., introduced a touch-enabled AI home art installation based on openvela. The product looks like a physical painting hung on the wall, where each item in the painting corresponds to a real-life service for the user, such as refrigerator ingredient management, calling friends, window weather display, and meal pairing decisions.
"We aim to provide users with emotional value and living convenience by combining AI with hardware," Xu Shengqi told reporters. The product's system is uniformly organized by a main Agent for memory, tasks, and tools. Users only need to say to the picture frame, "I want something light tonight," and the characters in the painting can catch the demand, provide suggestions and reasons, and after user confirmation, transfer the order to the mobile phone. Currently, the product is advancing debugging and prototype development based on Gemini-S1.
There is a long way to go from an artistic concept to hardware prototype validation and then to mass production. "At present, AI hardware prototypes are emerging in an endless stream, but mature products that can move towards mass production while taking into account stable operation, low power consumption, and user experience remain limited," pointed out Du Chao, head of openvela open source. From functional demonstration to actual use, the problems developers need to solve run through underlying adaptation and application design.
Reducing repeated investments in basic links is a path explored by the industry. Zhou Juejia, director of the Office of the Technology Committee of Xiaomi Group, stated that AI hardware development involves chip drivers, inference frameworks, model compression, power consumption optimization, and software upgrades. openvela aims to consolidate common capabilities into open infrastructure, allowing developers to focus on product innovation. Currently, openvela provides service frameworks for connectivity, graphics, and multimedia, as well as corresponding development and debugging tools. The reuse of these capabilities provides a common development foundation for different teams.
How to divide labor between devices and the cloud is also related to development efficiency. Liang Chaozhong, openvela system architect, introduced that the system provides a lightweight inference framework on the device side, while calling AI capabilities on the device, edge, and cloud through a unified interface, simplifying application development. This design provides a common entry point for different computing deployment methods, helping to reduce the handling of underlying differences during application development.
In development practice, the experiences of product developers also confirm the above views. "As the device-side system foundation, openvela, combined with Xiaomi's strong supply chain and model capabilities, can quickly adapt to development needs and support rich connection protocols. Developers can get started quickly, lowering the access threshold. In addition, it is convenient to access Xiaomi's existing ecosystem to achieve data interoperability between software and hardware," said Xu Shengqi.
Kingsley Mbewe, a master's student in computer technology at Beijing University of Posts and Telecommunications, developed a multi-pet smart feeder prototype, VelaPaw, based on openvela. During development, he found that the amount of code he didn't have to write himself was even more than he imagined—openvela performs like a complete system, measurable and patchable, with the vast majority of the code written by AI.
Meeting Usage Needs Gives AI Hardware Its Value
Technology choices must ultimately fall on usage needs. "A user-friendly AI product should help users use AI more simply and stably." Discussing the human-computer interaction experience, Qian Li, founder of Beijing Zhongjun Technology Co., Ltd., pointed out that in the voice assistant explored and developed by his team, users can record ideas at any time while walking or when it is inconvenient to operate the mobile phone, and then the cloud processes work information, accounting information, etc., making AI a more caring life assistant.
When users accept new devices, there is a learning cost and a process of accepting the new device. Faced with tasks that mobile phones can already complete, new devices need to answer what convenience they bring. For example, shortening operation steps, reducing interruptions during use, or adapting to environments where hands are inconvenient to operate can all be starting points for product design. In addition to whether functions are rich, whether users can use these functions naturally and smoothly also affects product value.
Between common capabilities and customized development, collaboration among industry partners is required. Lei Junbo, head of the Xiaomi Vela Business Unit of ThunderSoft, introduced that the company provides technical services around chip adaptation, system tailoring and optimization, and application development, and has participated in the mass production development of five watch products. The open-source system provides a common foundation, while solution providers complete the adaptation by combining specific devices and customer needs, promoting the integration of technology into products.
For AI hardware to become practical, it requires both more convenient development conditions and continuous product refinement. Making ideas easier to verify, and then consolidating the experience from verification into reusable capabilities, the industry can continuously turn imaginative and bold "inspirations" into user-friendly and durable "products," promoting the prosperity of the AI terminal industry.
Authors | Qi Xu, Hu Xiuqing Editors | Zhang Xinyi Art Editor | Maliya Supervisor | Lian Xiaodong