MiMo comes in a 7B parameter size, outperforming the o1-mini and QwQ-32B-Preview in the AIME 2024-2025 testing and LiveCodeBench v5.
The initial set of models released by Xiaomi includes the MiMo-7B-Base as the foundational model, MiMo-7B-RL-Zero trained with Reinforce from the base model, MiMo-7B-SFT trained in a Supervised manner, and MiMo-7B-RL trained from the SFT model. More information can be found on GitHub and Hugging Face.
Source: South China Morning Post
TLDR: MiMo models excel in testing and training methods, with various versions released by Xiaomi. Visit GitHub and Hugging Face for more details.
Leave a Comment