Content Extraction (OCR) 75 70 65 60 55 50 average score 338 70.1 66.5 63.3 63.1 58.7 6,552 66.2 1,973 63.8 790 55.0 7,593 54.2 NuExtract3 Chandra OCR 2 LightOn OCR 2 GLM-OCR olmOCR 2 Qwen3.5 4B GLM 4.6V Flash Gemma 4 4B Ministral 3 3B specialist models 338 # thinking tokens non-thinking score generalist models