MiMo is a language model project by Xiaomi focused on mathematical and general reasoning. The project covers the full training pipeline from pretraining through posttraining, including supervised fine-tuning (SFT) scaled to 6M instances and reinforcement learning (RL) training with extended context windows up to 48K tokens. Models are publicly released on HuggingFace and ModelScope, demonstrating competitive performance on benchmarks like MATH500 and AIME 2024.