SmolVLM: Redefining small and efficient multimodal models

Benchmark Model Rank Results
video-question-answering-on-tempcompass-overallSmolVLM 2.2B#16Accuracy: 53.7
zero-shot-video-question-answer-on-video-mme-w-o-subsSmolVLM 2.2B#52Accuracy (%): 52.1
video-question-answering-on-mlvuSmolVLM 2.2B#54Accuracy: 55.2
optical-character-recognition-ocr-on-ocrbenchSmolVLM 2.2B#60Score: 729
visual-question-answering-on-docvqa-testSmolVLM 2.2B#72ANLS: 0.800
video-question-answering-on-mvbenchSmolVLM 2.2B#80Avg.: 46.3
vision-language-understanding-on-mmstarSmolVLM 2.2B#95Accuracy: 46.0
chart-question-answering-on-chartqaSmolVLM 2.2B#961:1 Accuracy: 68.7
multimodal-reasoning-on-mathvistaSmolVLM 2.2B#105Accuracy: 51.5
multimodal-reasoning-on-mmmuSmolVLM 2.2B#120Accuracy: 42.0