Marker 2 is a complete rewrite of the open source document conversion pipeline from Datalab. It turns PDF, image, PPTX, DOCX, XLSX, HTML and EPUB files into markdown, JSON, HTML or chunks. The new version is built around Surya OCR 2, a 20 million parameter fast layout model and a rebuilt pdftext component that is three times faster than before.
Marker 2 offers three conversion modes. Balanced mode uses the Surya vision language model for layout and re‑OCR pages when the embedded text is poor; it scores 76.0 percent on the olmOCR‑bench benchmark and sustains about 2.9 pages per second on a single B200 GPU. Fast mode replaces the heavy VLM with a lightweight rf‑detr/onnx layout detector and minimal VLM use; it scores 66.6 percent and runs much cheaper. The –disable_ocr flag turns off all VLM calls, runs purely on CPU, needs no inference server and still extracts structure with the 20M layout model, delivering 43.6 percent at 23.7 pages per second.
Compared with MinerU’s pipeline backend, Marker balanced leads 76.0 to 72.7 percent overall and is more than five times faster at 2.9 versus 0.54 pages per second. Against Docling, Marker balanced wins 76.0 to 50.3 percent and is quicker at 2.9 versus 2.1 pages per second. LiteParse trades structure for speed, reaching over 1700 pages per second with OCR off but scoring only about twenty percent versus Marker’s CPU‑only 43.6 percent.
Licensing separates the tools. Marker’s code is Apache 2.0 while its model weights use a modified AI Pubs OpenRAIL‑M license that is free for research, personal use and startups under five million dollars in funding or revenue; beyond that a paid license is required. MinerU remains free up to 100 million monthly active users or twenty million dollars monthly revenue, Docling is MIT licensed, and LiteParse is open source with a paid cloud path for hard documents.
Choosing the right mode depends on your documents, hardware and budget. For high‑quality born‑digital PDFs on a GPU, use balanced mode. When cost matters and math is not critical, fast mode gives strong throughput. For air‑gapped or CPU‑only environments, the –disable_ocr option provides fast text‑layer extraction without any GPU or server. Always run the provided benchmark harness on your own corpus to confirm which tool fits your workflow.
#AI #Product #DocumentProcessing #OCR #ML #OpenSource