Xiaomi recently unveiled a groundbreaking AI hardware prototype called the Xiaomi AI Cube, a compact desktop system designed to run massive large language models (LLMs) completely locally, without reliance on cloud infrastructure. This mini PC uniquely integrates three self-designed XRING chips, showcasing the company’s ambitious shift from relying on Qualcomm processors to developing its own cutting-edge silicon.
The three proprietary chips powering the AI Cube are:
- XRING O3: A flagship AI system-on-chip (SoC) built on TSMC’s 3nm process with 24 billion transistors. It features a 10-core CPU, a 16-core G2 Ultra NX GPU, and a neural processing unit (NPU) delivering 200 tera operations per second (TOPS). Notably, this SoC supports LPDDR6 memory and will debut commercially in the upcoming Xiaomi 18 Fold phone. It is claimed to be the fastest mobile chip in the world based on benchmarks like AnTuTu.
- XRING O100: A high-bandwidth AI accelerator fabricated at 6nm using wafer-on-wafer 3D stacking technology. This chip delivers a staggering 1.22 terabytes per second (TB/s) memory bandwidth-approximately 16 times greater than current flagship smartphones-and can perform on-device inference at 330 tokens per second. It is designed for running large AI models locally and aims to overcome memory bottlenecks typical in AI hardware.
- XRING D100: China’s first 3nm smart driving AI chip featuring a 20-core CPU, a 16-core NPU, and up to 160 GB of unified memory. It is capable of running extremely large AI models with up to 200 billion parameters, supporting applications in autonomous driving.
The AI Cube itself sports an aerospace-grade aluminum chassis with 33,874 CNC-cut holes engineered for sophisticated cooling. It can sustain continuous power consumption up to 150 watts, allowing it to concurrently run two models-one with 120 billion parameters and another smaller 3 billion parameter model-with seamless switching between fast and slow modes depending on workload requirements. The system offers broad flexibility, enabling users to deploy any AI model locally, including models like DeepSeek or Alibaba’s Qwen series.
While the AI Cube is currently a prototype with no announced commercial release date (with some chips anticipated for 2027), it represents a serious strategic move by Xiaomi to establish dominance in the emerging local AI computing and edge inference market. Its combination of proprietary hardware with robust on-device AI capabilities challenges established players like Nvidia and Qualcomm, potentially disrupting the AI hardware landscape.
Beyond the AI Cube, Xiaomi showcased their wider silicon ambitions, including the adoption of their XRING chip family across different domains:
- Phones: The XRING O3 SoC is tailored for mobile devices, bringing cutting-edge AI performance and efficient memory support.
- AI edge computing: The XRING O100 facilitates running large-scale models with bandwidth efficiency on edge devices.
- Autonomous vehicles: The XRING D100 combines high CPU and NPU core counts with large unified memory pools to handle demanding smart-driving AI tasks.
This advancement aligns with broader trends in Chinese AI hardware development, emphasizing open weights, local inference without cloud dependency, and the integration of AI computing across consumer electronics and industrial applications. Xiaomi’s move exemplifies a growing ecosystem aiming to reduce latency, improve data privacy, and assert technological sovereignty through self-designed AI silicon.
In related AI developments, companies like Alibaba continue expanding their advanced multimodal models such as the Qwen series, capable of processing massive context windows and multiple data types (text, image, audio, video) for complex reasoning and agent workflows. Concurrently, platforms like B.AI provide developer-friendly environments offering free access to powerful multimodal AI models, lowering barriers for experimentation and integration.
Overall, Xiaomi’s AI Cube prototype signals an exciting step toward democratizing powerful AI computing at the edge, combining high-throughput hardware, energy-efficient design, and multi-model deployment flexibility. If commercialized successfully, it could redefine expectations for personal AI hardware, support diverse AI applications locally, and foster a competitive alternative to dominant cloud-based AI solutions.