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.
Sharing MiMo, a fascinating new series of open-source (Apache 2.0) language models from the Xiaomi MiMo team, specifically designed and "born for reasoning."
Their core idea is unlocking top-tier reasoning by optimizing pre-training for it (using reasoning-dense data), not just relying on post-training alignment like RL.
The results look impressive, their final RL-tuned 7B model reportedly matches OpenAI's powerful o1-mini on key math and code benchmarks!
Xiaomi has open-sourced the whole MiMo series – Base, SFT, and RL versions – on Hugging Face.
