- Published on
Powered by Self-Developed Chips with On-Device Xiaomi MiMo Model
- Authors
- Name
- DeepHorus
On August 24th, Xiaomi held the Xuanjie Chip Technology Communication Conference in the afternoon, officially launching three self-developed chips: the Xuanjie O3, O100, and D100. It also publicly demonstrated two prototype devices equipped with these chips for the first time. Xiaomi Group President Lu Weibing introduced one of the prototypes, named the Xuanjie O100, via social media. This device is specifically designed for on-device large model applications, aiming to provide high-speed AI inference capabilities.
The Xuanjie O100 prototype features a dual-chip collaborative computing architecture using the Xuanjie O3 and O100. To meet the demands of high-performance AI computing, the prototype's internal structure was completely redesigned, eliminating the traditional phone camera module and reworking the motherboard. It also incorporates an active air-cooling system capable of supporting up to 10 watts of heat dissipation to ensure stable chip operation under high loads. According to official data, the device is equipped with Xiaomi's self-developed Xiaomi MiMo on-device model, with measured inference speeds reaching 295 tokens per second.
Another prototype showcased is the Xiaomi AI Cube 'Prototype' mini host. This device features an aerospace-grade aluminum unibody, with over 33,874 CNC-precision-machined heat dissipation holes on the surface, ensuring a sustained high-performance release of 150 watts. Internally, the AI Cube integrates the Xuanjie O3, O100, and D100 chips and is equipped with 80GB of unified high-capacity memory. It can simultaneously deploy a 120B-parameter large model and a 3B-parameter on-device model, enabling intelligent switching between fast and slow systems.
The core of these two prototypes lies in demonstrating the powerful capabilities of Xiaomi's Xuanjie series chips in on-device AI computing. The high-speed inference capabilities of the Xuanjie O100 prototype and the massive model-carrying capacity of the AI Cube collectively point to a trend: future complex AI tasks will increasingly be completed on local devices, reducing reliance on the cloud, thereby improving response speed and protecting user privacy. Lu Weibing emphasized that these prototypes are high-speed AI demonstration terminals built for on-device large models.
Currently, both devices are still in the prototype stage. Xiaomi officially demonstrated their performance through several demo cases at the event but has not yet announced a specific commercialization timeline. Industry analysis suggests that Xiaomi's display marks its deep involvement in the AI chip and on-device large model fields. The deep integration of self-developed chips and terminal devices is expected to provide more robust underlying computing power support for its future smart hardware ecosystem. This technology communication conference undoubtedly offers an important window for the industry to observe Xiaomi's AI technology pathway.
