Xiaomi has just introduced Xiaomi Mimo, its first open-source AI large language model, designed for reasoning tasks.
Xiaomi MiMo
Xiaomi MiMo is a 7-billion-parameter model that excels in mathematical reasoning and code generation. It was developed by the company’s newly-formed Xiaomi Big Model Core Team and is said to match the performance of larger models like OpenAI’s 01-mini and Alibaba’s Qwen-32B-Preview.

The company believes that the reasoning effectiveness of MiMo, despite it being a smaller model, is driven by the base model’s potential, enabled through focused pre- and post-training strategies.
Pre-training
The reasoning ability of Xiaomi Mimo is built on an optimized pre-training process that includes improved data preprocessing pipelinne, enhanced text extraction tools, and multi-layered filtering. This was done to incrase the density of reasoning patterns.
The team compiled a dataset of 200 billion reasoning tokens and applied a three-stage data mixture strategy. The model was trained on 25 trillion tokens over three progressive training phases.
Post-Training Process
Meanwhile, the post-training stage used 130,000 mathematics and coding problems. The team also implemented a Test Difficulty Driven Reward system to address sparse rewards in complex tasks. There’s also Easy Data Re-Sampling for stable RL training on easier problems.
There’s also a Seamless Rollout Engine that cuts down GPU downtime, which increases training speed by 2.29x and a 1.96x boost in validation.
Xiaomi MiMo-7B comes in four versions:
- MiMo-7B-Base – the base model with strong reasoning potential
- MiMo-7B-RL-Zero – RL Model trained from the base
- MiMo-7B-SFT – a supervised fine-tuned model
- MiMo-7B-RL – RL model training from SFT
The MiMo-7B model series is open-source and accessible via Hugging Face. If you’re interested in the full technical report, you can check it out at GitHub.

