Datalab's Marker 2 Outperforms Competitors in Document Conversion

Datalab released Marker 2, a comprehensive rewrite of its open-source document conversion pipeline, on July 21, 2026. This new version converts various file formats including PDF, images, PPTX, DOCX, XLSX, HTML, and EPUB into markdown, JSON, HTML, or text chunks. The development team rebuilt the pipeline around three key components: Surya OCR 2, a fast 20 million-parameter layout model, and an enhanced pdftext module that is 3 times faster than its predecessor.

In performance benchmarks, Marker 2 achieved a balanced mode score of 76.0% overall on the olmOCR-bench, a third-party benchmark from Allen AI, with a score of 83.5% specifically on born-digital PDFs. On a single B200 GPU, Marker 2 sustains a throughput of 2.9 pages per second. This performance metric represents over a 5x increase in throughput compared to MinerU's pipeline backend, which scores 72.7% at 0.54 pages per second. For comparison, Docling scores 50.3% at 2.1 pages per second on the same benchmark harness. LiteParse's specific benchmark results were not detailed in this announcement but are implied to be lower than Marker 2.

Marker 2 offers three distinct conversion paths to cater to different needs. The 'balanced' mode, which utilizes the Surya VLM for layout analysis and re-OCR's pages with poor embedded text, delivers the highest quality, especially on GPUs, achieving the 76.0% olmOCR-bench score. The 'fast' mode employs a lightweight rf-detr/onnx layout detector and pdftext with minimal VLM usage, resulting in a 66.6% score and significantly lower costs. The '--disable_ocr' option provides pure text-layer extraction, running entirely on CPU without VLM calls and achieving a 43.6% score at 23.7 pages per second.

Structural changes in Marker 2 include device-aware mode selection by default, optimizing for GPU, CPU, or MPS based on the mode. Full CPU support is a significant enhancement, with the 'fast --disable_ocr' mode requiring no GPU or inference server, and the 20M layout model capable of reading columns, tables, and headers on CPU. The architectural shift involves multiple thin CPU workers sharing a single Surya inference server, with the parent process managing VLM concurrency across these workers, contributing to the substantial throughput improvements.