MiniCPM-o 2.6
OpenBMB · 2025-01-13 · 8.7B parameters
MiniCPM-o 2.6 is OpenBMB's 8.7B end-to-end omni model integrating SigLIP-400M, Whisper-medium, a speech synthesis component, and Qwen2.5-7B into a unified architecture, released in January 2025. It handles real-time speech conversation, vision understanding, and multimodal live streaming — accepting continuous video and audio streams and enabling voice cloning — while matching or exceeding GPT-4o on several vision and streaming benchmarks.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 86.1 |
| Capability | 121.1 |
| HallusionBench (OpenVLM) | 51.1 |
| MATH-Vision | 21.7 |
| MathVista (OpenVLM) | 73.3 |
| MM-Vet (OpenVLM) | 67.2 |
| MMMU Validation (OpenVLM) | 50.9 |
| MMStar (OpenVLM) | 63.3 |
| Video-MME | 63.9 |