Datalab Launches Marker 2 With Faster Document Conversion

Document conversion sits at the base of AI search, retrieval and training pipelines, where extraction quality and pages processed per second shape both model performance and computing costs. Datalab’s open-source Marker converts PDFs, images and Office files into Markdown, JSON or HTML. Marker 2 is a full rewrite combining selective OCR, document-layout detection and vision-language model repair, allowing the software to run across CPU-only systems and GPU deployments while choosing a default speed-accuracy mode by device.

Datalab released Marker 2 on July 20, 2026, integrating the 650-million-parameter Surya OCR 2 model, a 20-million-parameter layout detector and a rebuilt pdftext dependency that it says is three times faster. In Datalab’s tests on Allen Institute for AI’s 1,403-page olmOCR-bench, Marker’s balanced mode scored 76.0% and processed 2.9 pages a second on an Nvidia B200. MinerU’s pipeline backend scored 72.7% at 0.54 pages a second, giving Marker roughly 5.4 times the throughput.