MiniCPM-V 4.5 is designed as an efficient multimodal large language model that targets both single-image and multi-image understanding alongside high-frame-rate video analysis. Its architecture combines a Qwen3-8B language backbone with a SigLIP2-400M vision encoder, totaling roughly 8B parameters, and introduces a unified 3D-Resampler that compresses up to six 448x448 video frames into just 64 tokens. This 96x compression rate lets the model ingest many more frames without inflating inference cost, making it well suited for long-video and high-FPS scenarios on local devices, including iPad deployments demonstrated by the team.
In benchmark comparisons, MiniCPM-V 4.5 reaches a 77.0 average score on OpenCompass and competitive results on Video-MME while requiring less inference time than peers like GLM-4.1V-9B-Thinking and MiMo-VL-7B-RL on 8xA100 setups. The model supports controllable fast and deep thinking modes, hybrid multimodal reinforcement learning, and strong OCR plus document parsing behavior built on the LLaVA-UHD architecture for high-resolution images with arbitrary aspect ratios. An ophthalmology study evaluating MiniCPM-V 4.5 on bilingual clinical vignette questions found that enabling reasoning raised its scores on both English and Chinese datasets, though the model's overall accuracy trailed the comparison systems, suggesting it is best positioned for educational or exploratory multimodal workloads where its compact footprint outweighs the need for top-tier clinical precision.
