小米开源 Xiaomi Miloco 2.0：基于 MiMo 大模型的全屋智能 AI 方案

Today, Xiaomi released and open-sourced its whole-home smart AI solution, Miloco 2.0. Based on the self-developed MiMo large model, it is integrated into the Mijia ecosystem in the form of an Agent, upgrading from rule-based triggers to proactive home intelligence capable of memory, person recognition, and task execution.

Today (June 18), Xiaomi pushed Xiaomi Miloco to version 2.0.

It’s been exactly seven months since the first release in November last year. Back then, Xiaomi launched Miloco 1.0 (Xiaomi Local Copilot), which was more of a “proof of concept” — based on the MiMo-VL-Miloco-7B on-device vision-language model, it turned Mijia cameras into “eyes that can see images,” combined with natural language rule definitions so you could say to your home, “Turn on the desk lamp when reading.” But to be honest, 1.0 had a pretty high entry threshold — you needed an NVIDIA 30-series or higher GPU, you had to understand Docker, and manually configure MCP — in reality, it was something only the developer community could fully play with.

The change in 2.0 can be summed up in one sentence: From “smart home that can write rules using large models” to “a housekeeper with its own brain.”

No longer “if-then,” but “it figures it out itself”

The most exhausting thing about traditional smart homes is that you have to think through everything for the system — when to turn on the light, under what conditions to trigger, which devices link to which. Mijia, HA (Home Assistant), and Apple Home are all the same in essence — a set of finely tuned “condition triggers.” The more rules, the messier, and in most homes, those piles of automation scenarios are basically left unmaintained after half a year.

Miloco 2.0 aims to eliminate this problem.

Xiaomi has defined five key capabilities for it: general common sense, identity recognition, household memory, household tasks, and proactive intelligence. The first four are the foundation, and the last — “proactive intelligence” — is the real selling point of this upgrade: letting the system get things done before the user even opens their mouth.

In the official demo video, there’s a detail: the male homeowner arrives home later than usual, and instead of the cold “welcome home” from Xiao Ai, it deduces “probably worked overtime” and proactively says “Working overtime is tough, get some rest early.” Another example: when it detects a new Mijia device in the home, it automatically recommends integration scenarios, rather than waiting for you to open the app and add each one manually.

This logic sounds mystical, but boiled down it’s just three things: having common sense, remembering, and being able to reason.

Four foundational capabilities: making the “housekeeper” role real

Let’s examine the core features listed by the official sources and unpack their underlying meaning:

General common sense: No longer relying on preset rules to recognize danger. If a child picks up a knife, or an elderly person falls, the system can directly judge and issue graded alerts. This is powered by MiMo multimodal model’s visual understanding and world knowledge — traditional smart homes would need a specially trained fall detection algorithm and another model for dangerous objects; now you just hand it over to a VLM and it’s done.
Identity recognition: Face + posture fusion, done by the large model to recognize household members, and supports actively registering new members. This means the system can distinguish “Dad came home” from “The child came home,” thereby triggering different personalized actions. The importance here is that it turns “home” from a spatial concept into a collection of people with identities.
Household memory: Long-term accumulation of preferences and habits. For example, the system observes that you dim the lights, turn off the TV, and set the air conditioner to 26°C before bed every night. After some time, it will proactively ask, “Do you want to upgrade this routine into an automatic task?” This is basically adding a “long-term memory” module to the intelligent agent, similar to the role of a Memory system in an Agent framework within the large-model domain.
Household tasks: Upgrading from single-line rules to complex tasks that can operate long-term. “Remind to take medicine daily” or “track half an hour of exercise each day” — tasks that previously relied on separate apps — are now unified and abstracted into the Agent’s long-term tasks.

These four elements together support what is called “proactive intelligence” — the system observes, judges, and intervenes like a housekeeper.

Much lower deployment threshold — Mac mini can run it

The real deterrent in the 1.0 era was the hardware. Back then you needed x64 + a 30-series or better NVIDIA card + at least 16GB VRAM, which most households simply didn’t have running permanently.

2.0 directly changes the recommended configuration to a Mac mini:

Memory ≥ 4GB
Storage ≥ 256GB
24/7 permanent operation
macOS / Linux (Windows requires running under WSL)
A Xiaomi account + devices already connected to Mijia
Multimodal large-model API Key (MiMo-v2.5 for perception, MiMo-v2.5-pro for Agent)

This essentially means: A 4,000 RMB Mac mini can revitalize your whole-home smart setup. This is the biggest engineering improvement from 1.0 to 2.0 — not that the model itself is stunning, but that it’s finally “usable by ordinary enthusiasts.”

Another quietly released signal: Miloco 2.0 is integrated as an Agent plugin. This means it’s no longer a standalone closed system, but a component that can be embedded into a larger Agent ecosystem. If you’re already using Home Assistant or have built your own LLM workflow, you can theoretically attach Miloco as a capability module.

MiMo v2.5: Xiaomi finally plays its big card

To understand why Miloco 2.0 could evolve from rule-triggered to Agent reasoning, you have to look back at the MiMo line.

In April 2025, Xiaomi open-sourced its first reasoning large model MiMo, with 7B parameters, outperforming o1-mini and QwQ-32B-Preview on AIME 24-25 and LiveCodeBench v5 — a surprising event in the industry, marking the first time Xiaomi showed it was serious about large models. In November, Luo Fuli joined the MiMo team, significantly boosting its talent depth.

Miloco 2.0 now recommends MiMo-v2.5 for perception, and MiMo-v2.5-pro for Agent. The logic is clear: the perception layer handles high-frequency visual signal streams and needs to be lightweight and low-latency; the Agent layer deals with intent understanding, task planning, and tool invocation, requiring stronger reasoning ability. This “two-stage” division of labor has been the mainstream approach in open-source Agent frameworks over the past year.

It’s worth noting that the entire reasoning chain tightened privacy controls further: supports data isolation and one-click local data clearing. By principle, visual data does not leave the device. This is particularly important in the domestic smart-home context — remember that Mijia cameras are installed in bedrooms and living rooms.

What it really wants to do: give 1 billion IoT devices a brain

As of June last year, the number of IoT devices connected to Xiaomi’s AIoT platform (excluding phones, tablets, laptops) was already close to 1 billion. That’s a staggering number. Miloco 2.0’s real ambition isn’t to make a cool demo, but to give these 1 billion devices a unified brain.

This is where Xiaomi’s path differs most from Google and Apple:

Google Home uses “centralized cloud Gemini + device execution”
Apple bets on “on-device small models + strict ecosystem entry”
Xiaomi chooses “self-developed models + open source + edge deployment + third-party compatibility”

Open source is especially important here. The pain point of smart homes has never been that individual devices aren’t smart enough, but that cross-brand integration is too hard. Xiaomi, through its MCP protocol, has linked Mijia and Home Assistant; now in 2.0 it lowers the deployment threshold further and offers a web panel — essentially courting the geek community — to get them to install it at home, and in return contribute scenarios, rules, and data.

If this path really works, in the long term it could be more viable than a closed model like Apple HomeKit.

Some cold water worth pouring

After all these advantages, some sober judgment is needed.

First, the upper limit of “proactive intelligence” depends on reasoning quality. The demo’s “showing concern about overtime” is warm, but to get it right in every user’s home requires extremely accurate judgment of the current situation. Get it wrong once, and that’s the start of uninstallation. We won’t know the real-world misjudgment rate until the open-source community runs it for a while.

Second, Mac mini as the host has limited acceptance among domestic users. The recommended config says Mac mini, but many Mijia users don’t have a Mac at home. Linux deployment is supported, but still a barrier for ordinary users. In the short term, Miloco 2.0 is still a geek toy; to enter mainstream households, Xiaomi will need to release dedicated home AI host hardware — which is likely already in the works.

Third, the cost and latency of cloud Agent calls. MiMo-v2.5-pro runs via API, meaning each complex decision requires a cloud call. If the system really operates 24/7 according to the “proactive observation” design, the daily call volume won’t be small. If this can’t be greatly shifted to on-device processing, long-term usage costs will be problematic.

In conclusion

Miloco 1.0 was a “demo,” 2.0 is a “product.”

For the first time, it connects the often-mentioned “whole-home intelligence” with concepts like Agent, memory, and proactive intelligence, and lowers the engineering threshold to something an ordinary tech enthusiast can tinker with. The model is in Xiaomi’s hands, the framework is open-sourced on GitHub, and the hardware ecosystem has nearly 1 billion devices as its base — with these three elements in place, domestic smart-home manufacturers will likely be forced to respond in the second half of this year.

For developers, the most worthwhile thing now is to pull the code from GitHub and run it, to see how complex scenarios this Agent framework can truly handle. The smart-home track has been dormant for too long — it’s time someone redefined it.