Xiaomi Open-Sources Multimodal AI Model MiMo-VL-7B-2508
Xiaomi Open-Sources Advanced Multimodal AI Model
Xiaomi's AI research team has publicly released its MiMo-VL-7B-2508 multimodal large language model, marking a significant contribution to the open-source AI community. The release includes both Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) versions of the model.
Breakthrough Performance Metrics
The new model demonstrates exceptional capabilities across multiple domains:
- Subject reasoning: Achieved 70+ on MMMU benchmark
- Document understanding: Scored 94.4 on ChartQA
- Graphical interface positioning: Reached 92.5 on ScreenSpot-v2
- Video understanding: Improved to 70.8 on VideoMME
Technical Enhancements
The latest iteration shows substantial improvements in:
- Reinforcement learning stability
- Supervised fine-tuning processes
- Internal VLM Arena score (increased from 1093.9 to 1131.2)
User-Centric Features
The model introduces innovative interaction modes:
- Thinking mode: Displays complete reasoning chains (100% control success rate)
- Non-thinking mode: Direct answer generation (99.84% success rate with faster responses)
Users can toggle between modes using the
/no_thinkinstruction.
Available Model Versions
MiMo-VL-7B-RL-2508
- Recommended for general use cases
- Open source repository
MiMo-VL-7B-SFT-2508
- Suitable for custom fine-tuning
- Improved RL stability over previous versions
- Open source repository
Key Points
✅ New state-of-the-art performance in four core AI capabilities
✅ Dual-mode operation optimizes for accuracy or speed
✅ Fully open-sourced with commercial-friendly licensing
✅ Enhanced stability for reinforcement learning applications
