Xiaomi open-sourced its smart home AI framework Xiaomi Miloco (Mi Local Copilot) on GitHub in November 2025, which has already received 2.6k stars. Miloco uses the real-time video stream from Xiaomi home cameras as perceptual input, with a self-developed large language model at its decision-making core. Through the MiHome ecosystem, it connects all IoT devices in the home, aiming to allow users to define household rules and complex device interactions using natural language—for example, „automatically turn on the lights and play an announcement when someone enters the door“ or „notify parents when a child is detected alone in the living room.“ To protect privacy, visual understanding is completed on the local edge side. Xiaomi has also open-sourced the companion vision-language model MiMo-VL-Miloco-7B, specifically designed for home scene tasks, with the inference backend using llama.cpp. The entire system divides tasks into two phases: planning and visual understanding, and supports one-click deployment via Docker. Hardware requirements: x64 architecture, NVIDIA 30 series or higher graphics card, 8GB VRAM (12GB recommended), operating systems support Linux (Ubuntu 22.04+ recommended) and Windows 10+ (requires WSL2); macOS is not currently supported.
The project uses a non-commercial license. Without written authorization from Xiaomi, it may not be used for developing applications, web services, or other forms of commercial software, and is intended solely for research and personal exploration. Miloco’s core differentiator is transforming the camera from a mere security tool into a home-wide AI perception gateway, replacing cloud dependency with edge inference, enabling the home AI agent to respond in real time without uploading video data. Currently, the repository has 57 active Issues and 18 active Pull Requests, and the Xiaomi Miloco team is also openly recruiting externally.