---
title: "Xiaomi open-sourced MiMo, a 7B reasoning model built to run on the device, not the cloud"
date: 2026-11-14
category: Companies & Stack
site: NeuroAI
canonical: https://neuroai.site/a/na-corp-xiaomi-mimo
language: en
---

# Xiaomi open-sourced MiMo, a 7B reasoning model built to run on the device, not the cloud

> Xiaomi open-sourced its first reasoning model, MiMo-7B, and is folding it into phones, cars and smart-home gear.

Lei Jun's company spent years telling the world it was a hardware maker that happened to do software. Then it shipped a language model and gave it away.

Not a chatbot wrapper pulled from someone else's API. A reasoning model, small enough to sit on the silicon Xiaomi already builds.

The bet is that the model belongs next to the user — inside the gadget — instead of parked in a distant data centre that bills by the token.

## Key takeaways

- Xiaomi open-sourced Xiaomi MiMo (MiMo-7B), its first reasoning-focused large language model, on 30 April 2025 under an MIT licence; the company's 2025 interim report confirms the April release and a 7-billion-parameter design (Xiaomi IR, 2025 interim report; Securities Times, 30 April 2025).

- In May 2025 Xiaomi also released MiMo-VL-7B, an open-source multimodal model with 7B parameters, the interim report states (Xiaomi IR).

- On the AIME 24–25 maths set and the LiveCodeBench v5 coding benchmark, Xiaomi said its 7B MiMo-7B-RL scored above OpenAI's closed o1-mini and Alibaba's larger QwQ-32B-Preview; that claim comes from Xiaomi's own announcement and was repeated by Securities Times and Cailian Press (Xiaomi announcement; Securities Times; Cailian Press, 30 April 2025).

- Xiaomi's first-half 2025 R&D spend reached RMB 14.5 billion (≈ US$2.0B / HK$15.8B), up 35.8% year on year, with a record 22,641 R&D staff, per the interim report (Xiaomi IR).

- By December 2025 Xiaomi had released MiMo-V2-Flash, a mixture-of-experts model with 309B total parameters and 15B active, also open-sourced (Xiaomi MiMo official blog, December 2025).

- Xiaomi frames all of this under its "Human × Car × Home" (人车家) strategy, embedding models across phones, vehicles and smart-home devices (Xiaomi IR).

## A model that fits in a pocket

Most Chinese internet giants approached large models the way they approached search or shopping: build the biggest thing you can, host it centrally, and sell access. Xiaomi came at it from the other end. MiMo-7B is deliberately small — seven billion parameters — and the company's pitch is that efficient reasoning can run where the user already is.

The interim report is blunt about the motive. Xiaomi says it continued to advance "foundational large language models" and that MiMo was introduced as its first open-source reasoning model. The phrasing matters: open-source, not open-weights-only, and MIT-licensed, which lets commercial users fold it into products without negotiating a licence. Xiaomi's own product site spells out the intent plainly — on-device for phones, smart speakers and car cockpits, offline-capable and privacy-first, with cloud reserved for heavy lifting.

## The core team, formed in 2023

The model did not appear from nowhere. Securities Times reported that Xiaomi's large-model "Core" team was set up in 2023 at Lei Jun's push, focused on reasoning models and lightweight on-device deployment. The same report noted the group had grown into a sizable research organisation by 2025.

Xiaomi's own technical write-up says the 7B model was trained over three progressive stages on about 25 trillion tokens, with roughly 200 billion tokens of synthetic reasoning data, then refined with reinforcement learning. The company claimed its Seamless Rollout system sped training 2.29× and validation 1.96×. Those are Xiaomi's numbers; treat them as vendor benchmarks rather than independent tests.

What is verifiable without Xiaomi's word is the licensing and the release timing, both confirmed in the interim report filed with investors.

## The "Human × Car × Home" logic

Here is the part that is not about a leaderboard. Xiaomi is one of the few model builders that also ships the hardware the model runs on: 731.2 million global monthly active users across its ecosystem as of June 2025, and 989.1 million connected IoT devices on its AIoT platform (Xiaomi IR). A model that runs on a phone, a car cockpit or a smart speaker is a model that rides on top of a device base competitors have to buy their way into.

The interim report ties MiMo directly to that base: HyperOS integration for on-device responses, in-car assistants for the YU7, and the first-gen Xiaomi AI Glasses released in June 2025 with real-time translation across Chinese and ten other languages. The strategic read is that owning the model removes a dependency and keeps inference latency, privacy and cost inside Xiaomi's own stack.

According to Li An, chief scientist at BrainNet (脑机网), the interesting move is not the benchmark scores but that a device maker is treating the model as a component it owns outright, the way it already owns a display or a battery. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

## Where MiMo went next

The 7B release was a starting point, not a peak. The official MiMo blog shows the line moving fast: MiMo-Embodied in November 2025 for cross-domain embodied intelligence, then MiMo-V2-Flash in December 2025 — a MoE model with 309B total parameters and 15B active, a 256k context window, and open-sourced on Hugging Face and GitHub. The roadmap into 2026 lists a 1M-context Pro variant, a full-modal Omni model and a terminal coding agent.

That trajectory tells you the plan: a family, not a one-off, spanning text, vision, voice and embodied control — all pitched at running close to the user rather than in a centralised cloud.

## Honest limitations

This article relies on Xiaomi's own disclosures for the performance claims. The "beat o1-mini and QwQ-32B-Preview" result is Xiaomi's benchmark statement, repeated by Chinese financial press but not independently re-run here. The 2.29×/1.96× training-speed figures are vendor numbers. The R&D spend, user counts and IoT device counts come from Xiaomi's filed interim report and are the most solid figures in the piece. The "200 billion tokens of synthetic reasoning data" and the 2023 team-formation date come from Securities Times reporting rather than a primary filing, so treat them as well-sourced but secondary. I have not verified whether MiMo runs today in any specific mass-produced product beyond the AI Glasses and HyperOS mentions Xiaomi itself made, and on-device deployment at scale is a claim about intent as much as fact.

## What readers can do now

- If you build on open models, MiMo-7B and the V2-Flash MoE are MIT-licensed on Hugging Face and GitHub — check whether their licence and size fit an on-device or edge use case you already have.

- If you track Chinese AI strategy, watch whether Xiaomi keeps the "own the model, own the device" pattern or follows peers into larger centralised models; the V2-Flash scale-up is the signal to watch.

- If you are a developer or supplier, note that Xiaomi is recruiting into its large-model Core team and citing on-device and embodied deployment — a different hiring target from pure cloud-model labs.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-xiaomi-mimo
Free to quote with attribution and a link to the original.
