Xiaomi Launches Open-Source MiMo AI Model for Reasoning Tasks
April 30, 2025
Xiaomi
MiMo
AI
Large Language Model
Reasoning Tasks
Open-Source
Mathematical Reasoning
Code Generation
Reinforcement Learning
Benchmark Performance
Xiaomi has introduced MiMo, its first open-source AI large language model, optimized for reasoning tasks and rivaling larger models in performance.
Xiaomi MiMo AI Reasoning Model: Open-Source and Optimized for Reasoning Tasks
Xiaomi unveils MiMo open-source AI model for reasoning
Xiaomi has unveiled MiMo, its first open-source artificial intelligence large language model, specifically designed for reasoning tasks. Developed by the Xiaomi Big Model Core Team, MiMo is a 7-billion-parameter model that excels in mathematical reasoning and code generation, rivaling the performance of larger models like OpenAI’s o1-mini and Alibaba’s Qwen-32B-Preview.
Key Features of MiMo
- Optimized Pre-Training Process: MiMo’s reasoning ability is built on an enhanced pre-training process, which includes improved data preprocessing, advanced text extraction tools, and multi-layered filtering to increase reasoning pattern density. The model was trained on 25 trillion tokens over three progressive phases.
- Post-Training Reinforcement Learning: Xiaomi applied reinforcement learning using 130,000 mathematics and coding problems, verified by rule-based systems for accuracy and difficulty. Techniques like Test Difficulty Driven Reward and Easy Data Re-Sampling were used to stabilize training.
- Seamless Rollout Engine: This system reduces GPU downtime, delivering a 2.29× increase in training speed and a 1.96× boost in validation. It also supports Multiple-Token Prediction in vLLM, enhancing inference stability.
MiMo Model Variants
The MiMo-7B series includes four versions:
- MiMo-7B-Base: Base model with strong reasoning potential.
- MiMo-7B-RL-Zero: RL model trained from the base.
- MiMo-7B-SFT: Supervised fine-tuned model.
- MiMo-7B-RL: RL model trained from SFT, offering top-tier performance.
Benchmark Performance
MiMo-7B-RL has demonstrated strong performance across various evaluations:
- Mathematics: MATH-500: 95.8% Pass@1, AIME 2024: 68.2% Pass@1, AIME 2025: 55.4% Pass@1.
- Code: LiveCodeBench v5: 57.8% Pass@1, LiveCodeBench v6: 49.3% Pass@1.
- General: GPQA Diamond: 54.4% Pass@1, SuperGPQA: 40.5% Pass@1, DROP (3-shot F1): 78.7, MMLU-Pro (Exact Match): 58.6, IF-Eval (Prompt Strict): 61.0.
Availability
The MiMo-7B model series is open-source and accessible on Hugging Face. The full technical report and model checkpoints are also available on GitHub.
Sources
MiMo - Xiaomi's Open Source Model, Born for Reasoning
Open-source (Apache 2.0) LLM series 'born for reasoning.' Pre-trained & RL-tuned models (like the 7B) match o1-mini on math/code.
Xiaomi unveils MiMo open-source AI model for reasoning
Xiaomi has introduced MiMo, its first open-source artificial intelligence large language model, designed for reasoning tasks.
Xiaomi launches its first inference open-source large model Mimo ...
This model integrates pre-training and post-training to comprehensively enhance reasoning abilities. The AI competition both domestically and ...
Invite Gmail Contacts
