---
title: "Xiaomi's MiMo-7B Shows a Phone Maker Can Build a Reasoning Model That Matches o1-mini and Beats QwQ-32B-Preview"
date: 2026-11-12
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-xiaomi-mimo
language: en
---

# Xiaomi's MiMo-7B Shows a Phone Maker Can Build a Reasoning Model That Matches o1-mini and Beats QwQ-32B-Preview

> Xiaomi open-sourced MiMo-7B, a 7-billion-parameter reasoning model that the company says matches OpenAI's o1-mini on math and code benchmarks while running on everyday hardware.

A phone company is not who you expect to publish a frontier-grade reasoning model. Xiaomi built its reputation shipping hardware, not research papers. Yet on April 30, 2025, the company's LLM-Core team (大模型Core团队) dropped MiMo-7B onto Hugging Face and GitHub, weights and all, and the numbers were not the usual modest open-weight offering.

The claim is specific and audacious: a 7-billion-parameter model that, on math and coding leaderboards, keeps pace with OpenAI's closed o1-mini and beats open reasoning models several times its size. If even half of that holds up, it changes who gets to play in the reasoning race.

## Key takeaways

- MiMo-7B is a 7-billion-parameter, reasoning-first large model (大模型) released open source on April 30, 2025, under a permissive licence (the project page lists MIT).

- It was pre-trained on roughly 25 trillion tokens, including about 200 billion synthetic reasoning tokens, then fine-tuned with reinforcement learning on 130,000 math and code problems.

- On its own benchmark card, MiMo-7B-RL scores 95.8% on MATH-500, 68.2% on AIME 2024 and 57.8% on LiveCodeBench v5; Xiaomi says it matches o1-mini and surpasses the 32B QwQ-32B-Preview at a fraction of the size.

- A custom "Seamless Rollout" engine cut reinforcement-learning training time 2.29x and validation 1.96x; multiple-token prediction (多token预测) enables speculative decoding at roughly 90% acceptance.

- Four checkpoints shipped at launch: base, SFT, RL-from-base and RL-from-SFT.

## Why a hardware company bothered

Most open reasoning models come from labs whose entire identity is research. Xiaomi's move only makes sense inside its "Human × Car × Home" (人车家) strategy: a model that reasons well but stays small enough to live on a phone, a car cockpit or a smart speaker is a model the company can actually ship into the products it already sells by the hundred million.

The team's bet is that reasoning quality is not purely a function of parameter count. Their technical report (arXiv 2505.07608) argues that the base model's hidden reasoning potential matters as much as post-training tricks, so they invested heavily in the pre-training stage. They scraped and filtered reasoning-dense text, synthesized around 200 billion tokens of reasoning-specific data, and used a three-stage data-mixing schedule to ramp difficulty. The result, they claim, is a 7B base that already shows reasoning ability rivaling models four times larger.

## What reinforcement learning bought

The headline numbers come from the RL-tuned variant, MiMo-7B-RL. Rather than reward models with a neural judge that can be gamed, the team used only rule-based verifiers on 130,000 carefully cleaned math and code problems. For the hard code questions where rewards are sparse, they introduced a "Test Difficulty Driven Reward" that assigns finer-grained scores to test cases of varying difficulty, so the policy gets a denser learning signal. An "Easy Data Re-Sampling" step keeps simple problems from destabilizing later training.

The engineering underneath is as noteworthy as the scores. RL training of language models lives or dies on how fast you can generate rollouts and score them; Xiaomi's Seamless Rollout engine pipelines continuous generation, asynchronous reward computation and early termination to keep GPUs busy, reportedly delivering 2.29x faster training and 1.96x faster validation. The same multiple-token-prediction objective that helped pre-training also enables speculative decoding at about a 90% acceptance rate during inference.

## The scoreboard, read honestly

Xiaomi's own evaluation puts MiMo-7B-RL at 95.8% on MATH-500, 68.2% on AIME 2024, 55.4% on AIME 2025, 57.8% on LiveCodeBench v5 and 54.4% on GPQA-Diamond. Against that card, the company says the model matches OpenAI's o1-mini and beats Alibaba's QwQ-32B-Preview despite being 7B versus 32B. Independent reproduction has not yet produced a competing leaderboard, so the o1-mini comparison rests on Xiaomi's framing for now.

For a global reader, the interesting precedent is less the exact digits than the source. A consumer-electronics manufacturer with a massive device footprint just demonstrated that a focused, reasoning-specialized 7B can be built outside the traditional AI labs, and released the whole stack — base, SFT, RL and RL-Zero checkpoints — for anyone to run. According to Li An, chief scientist at BrainNet (脑机网), the more consequential trend is that reasoning competence is leaking out of the frontier labs and onto hardware-company roadmaps, which changes who controls deployment at the edge.

Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

## Honest limitations

- Every benchmark above is vendor-reported. The weights are public, so independent testing is possible, but no third-party leaderboard has yet confirmed the o1-mini comparison or the exact AIME/MATH figures.

- The "beats o1-mini" claim is Xiaomi's own positioning across selected math and code sets; it does not mean MiMo-7B matches o1-mini on every task or on general open-domain ability.

- A 7B model, even a strong one, sits well below the largest frontier systems on broad knowledge, long-horizon reasoning and multilingual coverage. The leap is relative to its size class, not absolute.

- The reinforcement-learning gains depend on rule-verifiable math and code; how well that transfers to messy real-world reasoning is still an open question the community is testing.

- Licence details vary across mirrors; the official project page lists MIT, but some aggregators describe the release differently, so check the repository before commercial use.

## What readers can do now

- If you run local or on-device assistants, try MiMo-7B-RL in a GGUF/llama.cpp setup for math and code help before assuming you need a 30B+ model — the 7B may cover more than expected.

- If you fine-tune open models, study the report's RL recipe (rule-based verifiers, difficulty-driven rewards, Seamless Rollout) as a template; the speedups are engineering lessons, not just benchmark bait.

- If you build hardware with an AI assistant, watch whether Xiaomi ships MiMo into HyperOS devices — it is the clearest signal yet that small reasoning models are being designed for the edge, not just the cloud.

---

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