Xiaomi MiMo: A New Open-Source Reasoning-Focused AI Model
Xiaomi’s Breakthrough in AI Reasoning
Photo by Digital Buggu on Pexels
Xiaomi Group has open-sourced its first large language model (LLM) designed specifically for reasoning—Xiaomi MiMo. This model enhances reasoning capabilities by integrating innovations from pre-training to post-training phases.
Following the announcement, Xiaomi-related stocks surged in the Hong Kong market. Kingsoft Cloud (03896.HK) jumped over 15%, while Xiaomi Group-W (01810.HK) and Kingsoft (03888.HK) rose more than 5%.
Superior Performance with Compact Size
Photo by Anete Lusina on Pexels
In benchmark tests, MiMo-7B outperformed larger models despite its smaller size. It surpassed OpenAI’s closed-source o1-mini and Alibaba’s QwQ-32B-Preview in mathematical reasoning (AIME 24-25) and coding competitions (LiveCodeBench v5).
With the growing trend of reinforcement learning (RL) collaboration sparked by DeepSeek-R1, models like DeepSeek-R1-Distill-7B and Qwen2.5-32B have become popular starting points.
Yet, MiMo-7B shows stronger potential in mathematical and coding tasks when trained on the same RL data.
Key Innovations Behind MiMo
Photo by Ahmet ÇÖTÜR on Pexels
The improvements in MiMo’s reasoning come from multiple breakthroughs in data, training, and algorithms:
Data
- Focused on high-quality reasoning datasets
- Generated ~200B tokens of synthetic reasoning data
Training
- Three-stage progressive difficulty training
- Total training volume: 25T tokens
Post-Training
- Efficient and stable RL algorithms
- Test Difficulty Driven Reward to address sparse rewards in complex tasks
- Easy Data Re-Sampling to stabilize RL training
Framework
- Seamless Rollout System speeds up RL training by 2.29x and validation by 1.96x
Open-Source Availability
MiMo-7B is now available on Hugging Face, with four models released: 🔗 GitHub Technical Report 🔗 Hugging Face Models
The Team Behind MiMo
MiMo is the first major project from Xiaomi’s newly formed Core LLM Team. While 2025 may seem like the latter stage of the AI race, the journey toward AGI remains long.
Xiaomi commits to practical innovation, pushing the boundaries of intelligence through bold exploration and creative problem-solving. Ask Mi Anything!
Keywords for Image Search:
Xiaomi MiMo, AI reasoning model, reinforcement learning, open-source LLM, Hugging Face, mathematical reasoning, coding benchmarks, AGI research
