MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

Benchmark Model Rank Results
visual-question-answering-vqa-on-mmeMiniCPM-V 4.5#2Score: 2500
visual-question-answering-vqa-on-mmt-benchMiniCPM-V 4.5#4Accuracy: 68.3
optical-character-recognition-ocr-on-omnidocbench-v1-0MiniCPM-V 4.5#5Edit Distance: 0.175
visual-question-answering-on-mm-vetMiniCPM-V 4.5#5GPT-4 score: 75.5
zero-shot-video-question-answer-on-longvideobenchMiniCPM-V 4.5#5Accuracy (% ): 63.9
optical-character-recognition-ocr-on-ocrbenchMiniCPM-V 4.5#10Score: 890
visual-question-answering-on-docvqa-testMiniCPM-V 4.5#13ANLS: 0.947
visual-question-answering-vqa-on-hallusionbench-average-accuracyMiniCPM-V 4.5#14Accuracy: 61.2
vision-language-understanding-on-mmbench-v1-1MiniCPM-V 4.5#15Accuracy: 84.2
video-question-answering-on-video-mme-with-subtitlesMiniCPM-V 4.5#17Accuracy: 73.5
chart-question-answering-on-chartqaMiniCPM-V 4.5#221:1 Accuracy: 87.4
video-question-answering-on-mlvuMiniCPM-V 4.5#23Accuracy: 75.1
multimodal-reasoning-on-mathvistaMiniCPM-V 4.5#31Accuracy: 79.9
vision-language-understanding-on-mmstarMiniCPM-V 4.5#33Accuracy: 72.1
video-question-answering-on-lvbenchMiniCPM-V 4.5#37Accuracy: 50.4
visual-question-answering-vqa-on-realworldqaMiniCPM-V 4.5#38Accuracy: 72.1
multimodal-reasoning-on-mmmuMiniCPM-V 4.5#42Accuracy: 67.7