After the official announcement of Luo Fuli, Xiaomi unleashes its first major AI move, enabling one-click connection of one billion IoT devices to the large model.
On November 14, Zhidx reported that just now, Xiaomi launched its first "large model + smart home" solution, Xiaomi Miloco, fully known as Xiaomi Local Copilot.
Screenshot of Miloco on GitHub. GitHub address: https://github.com/XiaoMi/xiaomi-miloco
Miloco uses Mi Home cameras as the source of visual information and its self-developed large language model, MiMo-VL-Miloco-7B as the core to connect all Internet of Things (IoT) devices in the home. The framework is open - source to everyone. The MiMo-VL-Miloco-7B model is optimized based on the MiMo model released by Xiaomi in April. Luo Fuli, the so - called "genius girl", recently joined the MiMo model team.
This is likely to be the "ChatGPT moment" for smart homes. As of June this year, the number of IoT devices (excluding smartphones, tablets, and laptops) connected to Xiaomi's AIoT platform has reached 989 million. Hundreds of millions of Mi Home cameras, Xiaoai speakers, table lamps, and other devices are expected to use large models.
From the Miloco page announced by Xiaomi, the main visual element of the page is a chat box similar to ChatGPT. There is a navigation bar for smart home devices on the left side of the chat box, including options such as AI Center, Model Management, MCP Service, and Device Management. "Camera Devices" is a separate column, displaying some videos recorded by smart cameras.
Display of the Miloco page
According to the design concept of Miloco, users can communicate with the smart home system. Through the inference and calculation of the large model, various intelligent needs and rules in family life can be automatically completed.
Miloco deploys its self - developed large model capabilities to edge devices in the home and provides an "AI brain" combined with the real - time visual information from Mi Home cameras. Specifically, Miloco has the following four major features:
1. New interaction paradigm: Based on the development paradigm of large language models, rule setting and complex device command control can be completed through natural language interaction.
2. New use of visual data: Using the camera data stream as the source of perception information, the large language model is used to analyze various home scene events contained in the visual data to respond to user queries.
3. Large language model on the device side: Home scene tasks are divided into two stages: planning and visual understanding. It uses the device - side model independently developed by Xiaomi to achieve device - side video understanding, ensuring family privacy and security.
4. Mi Home ecosystem: It is connected to the Mi Home ecosystem, supports the retrieval and execution of Mi Home devices and scenes, and supports sending customized content to push Mi Home notifications.
Miloco also encapsulates through the standardized MCP protocol to achieve the connection between the Mi Home ecosystem and the Home Assistant ecosystem, the world's largest open - source smart home community. At the same time, it supports the access of third - party IoT platforms.
From the software and hardware requirements announced by the project, the hardware requirements for deploying Miloco are not high. It only requires the hardware to be equipped with an x64 architecture, a graphics processor of NVIDIA 30 series or higher, and a storage capacity of 16GB or more.
Software and hardware requirements for Miloco deployment
It is reported that the differentiated experience of Miloco's whole - house intelligence relies on the support of the Xiaomi MiMo - VL - Miloco - 7B edge - side visual language large model and the four - layer complete architecture of "hardware - capabilities - applications - users".
The four - layer architecture of Miloco
The MiMo - VL - Miloco - 7B model makes its debut. It is optimized and built based on Xiaomi's self - developed MiMo - VL - 7B large model. With its visual - language fusion ability, it endows home cameras with the perception of "understanding the picture".
The MiMo - VL - 7B is developed through enhanced training of Xiaomi's first open - source inference large model, Xiaomi MiMo, released in April this year. In the public evaluation sets of mathematical reasoning (AIME 24 - 25) and code competitions (LiveCodeBench v5), with only 7B parameters, it scored higher than OpenAI's closed - source inference model o1 - mini and Alibaba's open - source inference model QwQ - 32B - Preview of Qwen. ("Xiaomi's First Inference Large Model Suddenly Goes Open - Source! Stock Price Rises Nearly 5%")
MiMo is the initial attempt of Xiaomi's large model Core team, which is full of talented people. On November 12, Luo Fuli, a former core member of DeepSeek, known as the "genius girl" in the industry, officially announced on her WeChat Moments that she had joined the Xiaomi MiMo team. ("Lei Jun Recruits a Former DeepSeek General! A Group Photo of the 40 - Member Large Model Team Is Exposed, Suspected to Enter Embodied Intelligence")
This time, in the acknowledgments of the Miloco project, nine original members of the Miloco team such as zhaoy and yangyongjie are mentioned, and thanks are given to the llama.cpp open - source project that provides the inference backend function.
Conclusion: Smart Homes May Welcome the "ChatGPT Moment", and the Battle Among Giants Is Imminent
Large models are accelerating their entry into the smart home scenario. Yesterday, Baidu just announced that tens of millions of sold Xiaodu devices will be upgraded to Super Xiaodu for free. Today, Xiaomi also takes a major step by announcing the implementation of large models in a powerful smart home form with Miloco.
In Xiaomi's vision, Miloco is expected to greatly simplify the original "mechanical and cumbersome" interaction process of smart homes. The experience bottleneck of traditional smart homes is expected to be broken, overcoming the dual constraints of "fixed pre - set rules" and "insufficient ecological collaboration".
While technology brings experience upgrades, the issue of data privacy becomes more severe. Xiaomi says its solution adheres to the principle of "privacy and security first". All visual data can be calculated on the home edge side without being transmitted to external servers, ensuring "no leakage of family privacy" from a technical perspective and dispelling users' concerns about data security.
This article is from the WeChat official account "Zhidx" (ID: zhidxcom). Author: Li Shuiqing, Editor: Yun Peng. It is published by 36Kr with authorization.