Xiaomi is a company best known for phones, TVs, and electric rice cookers. On April 30, 2025, it quietly became a serious player in open-source AI by releasing MiMo — a 7-billion-parameter model built specifically for reasoning, and built to run where Xiaomi's products actually live: on phones, in cars, and across the smart home.
A 7B model that punched above its weight
The headline result was deliberately provocative. Xiaomi said its reinforcement-learning–tuned MiMo-7B-RL, with just 7 billion parameters, outperformed OpenAI's closed o1-mini and Alibaba's larger QwQ-32B-Preview on public math-reasoning (AIME 24–25) and code-competition (LiveCodeBench v5) benchmarks. The point was not "we beat GPT" but "we beat much bigger models with a fraction of the size" — the classic Xiaomi "small-but-mighty" (以小博大) maneuver applied to AI.
That matters because parameter count is the enemy of on-device deployment. A 7B model can plausibly run on a phone or a car chip with careful optimization; a 32B or 70B model largely cannot. Xiaomi open-sourced four variants — Base, SFT, RL, and RL-Zero — and its official model site lists the MIT license, so commercial use is open. The technical report described improvements to the group-relative policy optimization (GRPO) algorithm, a test-difficulty–driven reward mechanism, and a self-developed training system that accelerated reinforcement learning by 2.29x.
From one model to a family
MiMo was not a one-off. The roadmap moved fast:
- November 21, 2025 — MiMo-Embodied released, targeting cross-domain embodied intelligence (the bridge from language model to robot).
- December 16–17, 2025 — MiMo-V2-Flash, a Mixture-of-Experts (MoE) flagship with 309B total parameters but only 15B active, a 56k context window, and 150 tokens/second inference.
- April 29, 2026 — MiMo-V2.5 series open-sourced (Pro with a 1M context window, Omni full-modal, and TTS), again under MIT.
- May 27, 2026 — API pricing permanently cut to US$1 per million input tokens and US$3 per million output tokens.
- June 10, 2026 — MiMo Code, an open-source terminal coding agent, launched.
The through-line is "reasoning-first, edge-friendly." Xiaomi explicitly markets a hybrid task-aware router that sends long-context or heavy-reasoning queries to the cloud and keeps light queries on-device — so a phone can answer offline in weak-signal areas and escalate to the cloud only when needed.
The market reaction
Investors noticed. Xiaomi's shares rose about 5.4% on April 30, 2025, the day of the MiMo announcement, with related Xiaomi ecosystem names (Kingsoft Cloud, Kingsoft) also jumping. The read was straightforward: a top phone vendor now has a credible, self-owned reasoning model, which de-risks its plan to embed AI across HyperOS, in-car agents, and smart-home orchestration without depending entirely on third-party models.
Why this matters globally
Two trends converge here. First, the "small model, big benchmark" race is real: if a 7B model can match a 32B model on the tasks that matter to users, the economic case for running AI locally — privacy, latency, offline use, no per-query cloud bill — gets much stronger. Second, device makers are becoming model makers. Apple, Samsung, Huawei, OPPO, vivo, and Xiaomi are all shipping on-device models now; Xiaomi's open-source stance is a different bet from Apple's closed one, and it gives global developers something they can actually fork and ship.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, Xiaomi's open-weight, edge-first approach reflects a broader Chinese strategy of making capable small models the default for consumer hardware, where cloud dependency is a liability rather than a feature.
What readers can do now
- If you build with models: pull MiMo-7B-Instruct from Hugging Face and benchmark it on your own reasoning and code tasks. At 7B, it is cheap enough to self-host on a single consumer GPU.
- If you buy phones or cars: expect "on-device AI" marketing to mean more soon. Ask vendors whether the model runs locally or just relays to a cloud — MiMo's hybrid router is the honest version of that story.
- If you follow the market: track MiMo's API price as a signal. A permanent US$1/US$3 per-million-token rate resets expectations for what "cheap inference" means in 2026.
Honest limitations
Benchmark claims (AIME, LiveCodeBench, SWE-bench) are Xiaomi's own from its technical report and model-site listings; they were not independently reproduced here. The license is listed as MIT on Xiaomi's official model site, though some secondary Chinese sources describe Apache 2.0 — the discrepancy was not resolved, so readers should verify the license on the source repository before commercial use. The ~5.4% share-price move is reported by Chinese financial media (Securities Times) and reflects a single-day reaction, not a durable valuation judgment. The "edge-first" strategic reading is the author's interpretation of publicly dated releases.