Datalab Marker 2 Outperforms Competitors in Document Conversion

Datalab released Marker 2 this week, a complete rewrite of its open-source document conversion pipeline. This updated system converts various file formats including PDF, images, PPTX, DOCX, XLSX, HTML, and EPUB into markdown, JSON, HTML, or text chunks. The development of Marker 2 was centered around three key components introduced in recent months: Surya OCR 2, a 20 million parameter fast layout model, and an enhanced pdftext module that is reportedly 3 times faster than its predecessor.

Performance benchmarks from olmOCR-bench, a third-party evaluation from Allen AI, indicate Marker 2's superior capabilities. In its balanced mode, Marker 2 achieved an overall score of 76.0% and an 83.5% score specifically on born-digital PDFs. This mode can process documents at a rate of 2.9 pages per second on a single B200 GPU, which is over 5 times the throughput of MinerU's pipeline backend. MinerU's backend achieved a score of 72.7% at a rate of 0.54 pages per second. Docling, another competitor, scored 50.3% at 2.1 pages per second on the same benchmark.

Marker 2 offers three distinct conversion paths to cater to different needs. The 'balanced' mode, utilizing the Surya VLM for layout analysis and re-OCR for embedded text, prioritizes quality and performs best on GPUs, achieving 76.0% on olmOCR-bench. The 'fast' mode employs a lightweight rf-detr/onnx layout detector and pdftext with minimal VLM usage, offering a lower cost at 66.6% accuracy. The '--disable_ocr' mode focuses solely on text-layer extraction, running entirely on CPU without VLM calls, achieving 43.6% accuracy and a throughput of 23.7 pages per second.

Architectural changes in Marker 2 contribute significantly to its throughput. The system now features multiple thin CPU workers sharing a single Surya inference server, allowing the parent process to manage VLM concurrency across workers. This design enables throughput to scale with server capacity rather than per-process VRAM. Datalab reports that the balanced mode can sustain approximately 2.9 pages per second, a substantial increase from a single-stream rate of around 0.3 pages per second on the same hardware. Users should note that Marker 2 requires Python 3.10+ and introduces breaking changes.