Paper
MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining
Authors
Xi Xia, Bowen Shen, Cici, Dawei Zhu, Di Zhang, Gang Wang, Hailin Zhang, Huaqiu Liu, Jiebao Xiao, Jinhao Dong, Liang Zhao, Peidian Li, Peng Wang, Shi-jie Yu, Shimao Chen, Weikun Wang, Wenhan Ma, Xia Deng, Yi Huang, Yi-Hao Song, Zi-Ang Jiang, Bowen Ye, Can Cai, Chenhong He, Dong Zhang, Duo Zhang, Guoan Wang, Hao Tian, Hao-Yun Zhao, Hengxu Qu, Hong-Mei Xu, Jun-Miao Shi, Kainan Bao, Qing Fang, Kang Zhou, Kang Zhou, Lei Li, Menghang Zhu, Nuo Chen, Qiantong Wang, Shao-yang Liu, Shicheng Li, Shuhao Gu, Shu-Qin Ren, Shuo Liu, Sirui Deng, Weiji Zhuang, Weiwei Lv, Wenyu Yang, Xin Zhang, Xing Yong, Xing Zhang, Xi-Na Song, Xin-Ming Xu, Xu Wang, Yihan Yan, Yunwu Tu, Yu-Shi Tian, Yudong Wang, Yue Yu, Zhenrui Lin, Zhichao Song, Zihao Yue
Abstract
We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.