MiniCPM-o 4.5
MiniCPM-o Omnimodal Series
MiniCPM-o 4.5 is the most capable omnimodal model in the MiniCPM-o series. It is built end-to-end on SigLIP2, Whisper-medium, CosyVoice2 and Qwen3-8B with 9B total parameters. The modality encoders / decoders and the LLM are densely connected via hidden states, which gives better information flow and control and lets training fully exploit rich multimodal knowledge.
Its signature capability is full-duplex multimodal live streaming: real-time video and audio input streams run alongside text and speech output streams without mutual blocking, so the model can see, listen and speak at the same time — and go beyond reactive answers to proactive interaction, such as initiating a reminder based on its continuous understanding of the scene. On vision it averages 77.6 on OpenCompass across 8 benchmarks, surpassing GPT-4o and Gemini 2.0 Pro and approaching Gemini 2.5 Flash with only 9B parameters, and supports both instruct and thinking modes in a single model. On speech it supports bilingual (English / Chinese) real-time conversation with configurable voices, plus voice cloning and role play from a single reference clip, outperforming CosyVoice2 on cloning. On OCR it handles images up to 1.8 million pixels in any aspect ratio and video up to 10fps, reaching state of the art for end-to-end English document parsing on OmniDocBench.
Highlights
- End-to-end omnimodal — vision encoder + speech encoder + LLM + TTS densely connected via hidden states
- Full-duplex live streaming — hear, see, and speak simultaneously
- Voice cloning and configurable assistant voice
- HuggingFace:
openbmb/MiniCPM-o-4_5· ModelScope:OpenBMB/MiniCPM-o-4_5
For real-time, full-duplex demos see the MiniCPM-o-Demo repository — it ships a turn-key web demo for the o-series with PyTorch + CUDA inference and a frontend.
Related Resources
| Topic | Link |
|---|---|
| Inference | MiniCPM-o 4.5 inference guide |
| Deployment | MiniCPM-o 4.5 deployment guide |
| Quantization | MiniCPM-o 4.5 quantization guide |
| Online demo | MiniCPM-o 4.5 live demo |
| Technical report | MiniCPM-o 4.5 Technical Report |