Xiaomi launched MiMo, its first open-source AI reasoning model series, featuring a compact 7B-parameter design that rivals larger models like OpenAI’s o1-mini. It excels in math and coding, scoring 55.4 on AIME, and is available on GitHub/Hugging Face.
On Tuesday, Xiaomi unveiled MiMo, its first open-source AI model series focused on advanced reasoning. The innovation lies in delivering high-performance reasoning within a compact framework, rivaling larger models like DeepSeek R1, Alibaba’s Qwen QwQ-32B, OpenAI’s o1, and Google’s Gemini 2.0 Flash Thinking. The MiMo family includes four specialized models, each tailored for distinct applications.
Solving the Size Challenge in Reasoning AI
Most reasoning AI models require at least 24 billion parameters to maintain strong coding and mathematical abilities—a hurdle for smaller models. Xiaomi’s breakthrough comes with MiMo-7B, a 7-billion-parameter model that reportedly matches OpenAI’s o1-mini and surpasses some 32B models in performance. The model was pre-trained on 25 trillion tokens, leveraging optimized data preprocessing, enhanced text extraction, and multidimensional filtering for efficiency.
Benchmark Performance
Internal tests highlight MiMo’s competitive edge:
- MiMo-7B-Base scored 75.2 on BIG-Bench Hard (BBH) for reasoning.
- MiMo-7B-RL-Zero, fine-tuned with zero-shot reinforcement learning, excels in math and coding, achieving 55.4 on AIME—4.7 points higher than o1-mini.
Open-Source Availability
Xiaomi has released MiMo on GitHub and Hugging Face, including a detailed technical paper covering its architecture and training processes. While it’s a text-only model (no multimodal support), its open-source nature allows developers to explore its capabilities—though dataset specifics remain undisclosed.
This launch positions Xiaomi as a growing player in AI, offering a smaller, yet powerful alternative to mainstream reasoning models.