A phone in Beijing finishes a tough math problem with no server ping. A pair of glasses in a Shanghai café translates a live conversation. A seven-billion-parameter model that a week earlier lived only inside one company is now sitting on GitHub.
The company is Xiaomi (小米), and the model is MiMo. Released April 30, 2025, Xiaomi MiMo is the phone maker's first open-source large model (大模型) built specifically for reasoning. Its arrival says something about where Xiaomi thinks AI actually lives: not in a distant data center, but in the device already in your hand.
Small model, loud claim
MiMo ships at 7 billion parameters — modest by the standards of trillion-parameter headline chasers. Xiaomi's pitch is efficiency over size. The company says MiMo-7B beat OpenAI's closed o1-mini and Alibaba's larger QwQ-32B-Preview on public math (AIME 24-25) and code (LiveCodeBench v5) benchmarks at the 7B scale. Those are Xiaomi's own benchmark claims, and they are the kind a rival lab would contest — but the underlying point is the one that matters for a hardware company: a model this small can run where the user is.
A month later, in May 2025, Xiaomi followed with MiMo-VL-7B, a multimodal version that adds vision to the reasoning core. For a firm whose products are cameras, screens, and sensors, a model that can read images is the bridge from "assistant" to "device that understands its surroundings."
The XiaoAI through-line
MiMo did not appear from nowhere. It extends a longer Xiaomi thread: 小爱同学 (XiaoAI), the voice assistant that has shipped on hundreds of millions of Xiaomi phones, TVs, and speakers, is being rebuilt as 超级小爱 (Super XiaoAI), an on-device (端侧) assistant that can handle more of the request locally. The strategic logic is consistent — keep the interaction on the hardware, not dependent on a round trip to a cloud you don't control.
That matters more than it sounds. A cloud assistant is a recurring cost and a privacy question. An on-device model is a feature you bought with the phone. For a company that sells margin-thin hardware, owning the model is owning the upgrade path.
Glasses, cars, and a homemade chip
The hardware story got concrete in mid-2025. Xiaomi's first AI glasses (第一代小米AI眼镜), released in June 2025, carry 超级小爱 and can translate in real time across Chinese and ten other languages — a live interpreter you wear. Separately, Xiaomi has pushed 小爱 voice, multimodal interaction, and voice car control into the Xiaomi YU7 cockpit, turning the car into another surface for the same assistant.
Underneath all of it is silicon. In May 2025 Xiaomi announced XRING O1 (玄戒O1), its first self-developed flagship processor — a second-generation 3nm chip with a 10-core CPU and 16-core GPU. A custom SoC is exactly what lets a 7B-class model run smoothly on a phone or glasses without melting the battery. The model and the chip are the same bet told twice.
Why a phone maker open-sources at all
Xiaomi is not a cloud vendor the way Tencent or Alibaba is. Giving MiMo away does not directly sell server time. So why publish weights? Three reasons fit the hardware playbook:
- Developer reach. Open weights let app makers build Xiaomi-shaped experiences, thickening the ecosystem around the devices.
- Talent signal. A public, benchmarked model is a recruiting and prestige marker in a market where AI engineers are scarce.
- On-device proof. Releasing a small, capable model demonstrates the exact thing Xiaomi wants inside its products — intelligence that fits in a pocket.
The contrast with the cloud giants is the point. Alibaba and Tencent open models to pull you into their clouds. Xiaomi opens a model to pull you into its devices.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the race to push large models onto phones and wearables is becoming the real differentiator for consumer-electronics firms, because cloud APIs are now a commodity.
What readers can do now
- If you build mobile or edge apps, prototype with MiMo-7B or MiMo-VL-7B — a 7B model is small enough to run on consumer hardware and cheap enough to ship.
- If you buy enterprise devices, ask vendors whether their "AI" is on-device (端侧) or cloud-backed; the answer changes your latency, cost, and data exposure.
- If you follow the China AI stack, watch whether more hardware makers (not just clouds) start open-sourcing small models — it signals where the next battlefront is.
Honest limitations
This article draws on Xiaomi's interim 2025 report (its investor-relations filing), Chinese tech press (Securities Times, Tencent News), and Xiaomi's model announcements. The benchmark claims that MiMo-7B surpassed OpenAI's o1-mini and Alibaba's QwQ-32B-Preview are Xiaomi's own statements, not results I verified independently; treat them as marketing claims pending third-party reproduction. Figures such as the AI team size (~3,000 people, attributed to Lei Jun in 2023) and the XRING O1 specifications come from company disclosures and may shift. I have not run MiMo on device myself; capability statements reflect vendor and launch reporting.