MiniCPM-V 4.5

MiniCPM-V 4.5 is an open-source, edge-side multimodal large model with 8 billion parameters, jointly developed by the Natural Language Processing Lab at Tsinghua University and FaceWall AI. It achieves a 96x video compression rate through its 3D-Resampler architecture, supports high-frame-rate video understanding at 10fps, and leads in benchmarks such as MotionBench. With an OpenCompass score of 77.2, it excels in OCR capabilities, supports over 30 languages, and enables seamless switching between fast and deep thinking modes.

Note: The MiniCPM-V 4.5 task allows for both video analysis and document analysis.

Introduction:https://zhuanlan.zhihu.com/p/1944351659238089720