MiMo

Open-source (Apache 2.0) LLM series 'born for reasoning.' Pre-trained & RL-tuned models (like the 7B) match o1-mini on math/code. Base/SFT/RL models released.

MiMo is an open-source language model series developed by Xiaomi, specifically optimized for reasoning tasks. Licensed under Apache 2.0, it includes pre-trained, supervised fine-tuned (SFT), and reinforcement learning (RL)-tuned models, with the 7B variant matching OpenAI's o1-mini in math and code performance.

The model excels due to its focus on reasoning from pre-training, leveraging enhanced data filtering, synthetic reasoning data, and a three-stage training approach. Post-training employs curated math and code problems with rule-based rewards to refine performance. MiMo also features infrastructure improvements, such as a seamless rollout engine, accelerating RL training.

Available models include the base version, SFT, and RL-tuned variants, all hosted on Hugging Face. MiMo demonstrates that smaller models can achieve reasoning capabilities comparable to larger ones, offering valuable insights for the AI community.