Xiaomi Unveils Open-Source MiMo-7B AI Model Series, Surpassing Key Rivals in Reasoning and Code Tasks

Chinese tech giant Xiaomi has made a significant leap in artificial intelligence by unveiling its first open-source large language model series, MiMo-7B. The family consists of four versions: "MiMo-7B Base," "MiMo-7B-SFT," "MiMo-7B-RL-Zero," and "MiMo-7B-RL." Among these, the RL (Reinforcement Learning) variant stands out for its consistency and effectiveness, as it’s built upon the SFT (Supervised Fine-Tuning) model.

According to Xiaomi's official statements, MiMo-7B was trained from scratch and demonstrates extraordinary reasoning capability. The RL version, in particular, excels in complex mathematical reasoning and code generation tasks — even outperforming OpenAI’s closed-source o1-mini model. Official results show that the MiMo-7B base model achieved a 75.2 score in the demanding BBH test, outperforming other open-source 7B models. The MiMo-7B-RL version marked 55.4 in the 2025 AIME benchmark, scoring 4.7 points higher than OpenAI’s acclaimed o1-mini.

Developed using 25 trillion tokens, the MiMo-7B series combines robust pretraining and state-of-the-art reinforcement learning techniques. All four versions are freely available for download via Hugging Face and ModelScope platforms, encouraging broad community participation.

MiMo-7B demonstrates remarkable performance in both mathematics and coding tasks. In the LiveCodeBench v5 coding benchmark, its RL model reached a success rate of 57.8%, significantly surpassing many competitors. With this open-source initiative, Xiaomi has made a notable contribution to the AI developer community and reinforced its position as a leading player in the evolving AI landscape.