Marker 2 is a complete rewrite of Datalab’s open‑source document conversion pipeline. It turns PDFs, images, PPTX, DOCX, XLSX, HTML and EPUB files into markdown, JSON, HTML or text chunks. The system is built around three core components: Surya OCR 2, a 20‑million‑parameter fast layout model, and a rebuilt pdftext engine that runs three times faster than before.
Marker 2 offers three conversion paths. Balanced mode uses the Surya vision‑language model for layout and re‑OCR when the embedded text is poor, delivering the highest quality at 76.0 % overall score on the olmOCR‑bench benchmark and 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 plus pdftext, scoring 66.6 % while achieving roughly 7.4 pages per second on GPU and far lower cost. The –disable_ocr path does pure text‑layer extraction, needs no GPU or inference server, runs entirely on CPU at 23.7 pages per second with a 43.6 % score.
Compared to MinerU’s pipeline backend, Marker 2 balanced is 5.4 × faster (2.9 vs 0.54 pages per second) while scoring higher (76.0 % vs 72.7 %). Against Docling, Marker 2 leads on both accuracy (76.0 % vs 50.3 %) and throughput (2.9 vs 2.1 pages per second). LiteParse trades structure for speed: it can reach over 1700 pages per second with OCR off but scores only about 20 % on CPU, whereas Marker 2’s CPU‑only –disable_ocr mode stays above 43 %.
Licensing differs across tools. Marker 2 code is Apache 2.0, but its model weights use a modified AI Pubs OpenRAIL‑M license that is free for research, personal use and startups under $5 M funding or revenue; beyond that a paid license is required. MinerU remains free up to 100 M monthly active users or $20 M monthly revenue. Docling is MIT‑licensed, and LiteParse is open source with a paid cloud option for hard documents.
For teams choosing a parser, consider document type, hardware availability, budget and licensing constraints. High‑volume born‑digital PDFs with a GPU benefit from balanced mode. Cost‑sensitive workloads that can skip math equations work well with fast mode. Air‑gapped or CPU‑only environments should use fast –disable_ocr or Docling. Scanned archives or math‑heavy papers may need a full‑page VLM such as Chandra 2. Node/TypeScript or browser stacks favor LiteParse, while mixed audio/email inputs point to Docling.
Key takeaways: Marker 2 balanced gives 76 % accuracy at 2.9 pg/s, over five times MinerU’s throughput; it beats Docling on speed and score; fast –disable_ocr runs pure CPU at 23.7 pg/s; licensing splits the field with Marker requiring paid weight use above $5 M; all numbers are reproducible with the supplied benchmark harness.
#AI #Product #DocumentConversion #OCR #MLTools #OpenSource