OCR'ing 100k pages with open-source VLMs on Modal