Xiaomi releases its first inference model MiMo-7B and announces full open source

According to Xiaomi's official account "Xiaomi MiMo", Xiaomi opened its first big model MiMo-7B for "Reasoning" today, surpassing OpenAI o1-mini and Ali Qwen-32B in AIME mathematical reasoning and LiveCodeBench programming evaluation. MiMo improves its reasoning ability by combining large-scale inference data pre training with reinforcement learning post training, proposing innovative algorithms such as Test Difficulty Driven Reward, and leading in reinforcement learning potential among similar 7B models. The entire series has been open sourced to Hugging Face, and related technical reports have been released simultaneously.