Xiaomi Releases MiMo Open-Source Reasoning Models, 7B Matches o1-mini on Math and Code
English summary
Xiaomi unveiled the MiMo series, an Apache 2.0 licensed open-source LLM family designed for reasoning. The models include pre-trained and RL-tuned variants, with the 7B version matching o1-mini performance on math and code benchmarks. Base, SFT, and RL model checkpoints have been publicly released.
Chinese summary
小米发布了 MiMo 系列，这是采用 Apache 2.0 许可的开源大语言模型家族，专为推理任务而生。模型包含预训练和强化学习调优版本，其中 7B 模型在数学和代码基准测试上的表现与 o1-mini 相当。基座模型、SFT 和 RL 模型权重均已公开发布。
Key points
Fully open-source under Apache 2.0 license
采用 Apache 2.0 完全开源许可
7B model matches o1-mini on math and code reasoning benchmarks
7B 模型在数学和代码推理基准上达到 o1-mini 水平
Pre-trained, SFT, and RL-tuned model variants are all released
发布了基座模型、监督微调（SFT）和强化学习（RL）调优的全套模型
Model series is specifically designed and tuned for reasoning tasks
模型系列专门针对推理任务进行设计和优化