Xiaomi introduces MiMo: the first open-source AI model focused on logical reasoning. It wants to beat OpenAI

Xiaomi is entering the world of large language models (LLMs) with its own open-source project. Their new model MiMo is the first large AI system from Xiaomi’s lab, developed by a dedicated team called “Xiaomi Big Model Core Team”.

Join the biggest Xiaomi group in Slovakia

Focused on mathematical reasoning and code generation

MiMo has a size of 7 billion parameters and focuses on areas like mathematical reasoning and code generation. What’s interesting is that, in terms of performance, it competes with much larger models such as OpenAI o1-mini (the creator of ChatGPT) or Alibaba Qwen-32B-Preview.

Achieving this level of computational intelligence with a “smaller” model is a challenge, because most existing RL models (reinforcement learning) have up to 32 billions of parameters. Xiaomi, however, believes that MiMo’s success is driven by a strong pretraining architecture and subsequent strategic fine-tuning. At the same time, MiMo’s smaller size makes it well-suited for deployment in enterprise applications or on devices with limited performance.

How did Xiaomi train MiMo? The foundation is a special pre-training

The key to MiMo’s high performance is a well-thought-out pre-training process. Xiaomi optimized the data processing pipeline, improved text-extraction tools, and implemented multi-level filters that increased the occurrence of logical reasoning formulas in the training data.

The model was trained on a dataset with more than 200 billion tokens, representing tasks from logic, mathematics, and programming. Xiaomi used Multiple-Token Prediction as the training objective, which helped not only improve accuracy but also speed up inference—the model’s actual “thinking” over new tasks.

Post-training: Reinforcement learning with an emphasis on accuracy

After pretraining, the team followed with a fine-tuning phase using reinforcement learning. MiMo was exposed to more than 130,000 mathematical and programming tasks, which were verified using rules for accuracy and difficulty. To ensure the model rewards results correctly even in complex situations, Xiaomi created its own Test Difficulty Driven Reward system.

A key step was deploying the Seamless Rollout Engine, delivering 2.29× faster training and 1.96× faster validation. This infrastructure also supports Multiple-Token Prediction and improves model stability during inference.

MiMo model variants (7B) and how they’re used

Xiaomi released four MiMo-7B variants at once, each with a different use case. They are:

• MiMo-7B-Base — a base model with high potential for reasoning
• MiMo-7B-RL-Zero — a model trained with reinforcement learning from the Base version
• MiMo-7B-SFT — a model fine-tuned via supervised learning
• MiMo-7B-RL — the final RL model from the SFT version, achieving comparable performance to OpenAI o1-mini

MiMo delivers top performance on benchmarks

The most advanced MiMo-7B-RL version achieves excellent results across a variety of standardized tests:

These figures show that even a mid-sized model can compete with much larger solutions—if it’s designed correctly and fine-tuned.

Availability and open-source approach

The entire MiMo-7B series is available as an open-source model on the Hugging Face platform. Xiaomi also published the full technical report and model checkpoints on GitHub, paving the way for developers and researchers who want to build their own solutions on top of MiMo.

To get more updates from the Xiaomi world, keep following our Xiaomi Planet website — join our Facebook group, where we build an active community and help each other out. Also subscribe to our Telegram, where we add the best offers and coupons from China every day.