Datalab's Marker 2 Rewrite Claims 5x Throughput Lead Over MinerU in Document Parsing

Datalab shipped Marker 2 on July 21, 2026, a ground-up rewrite of its open source document conversion pipeline, and the numbers behind it target a very specific pain point: turning messy PDFs into clean structured text without burning GPU hours.

The headline figure is throughput. On olmOCR-bench, a third-party benchmark from Allen AI built from 1,403 PDFs and roughly 8,400 pass/fail unit tests, Marker 2's balanced mode scores 76.0% overall and 83.5% on born-digital PDFs. MinerU's pipeline backend trails at 72.7% accuracy but processes only 0.54 pages per second, less than a fifth of Marker 2's pace in the same mode.

That speed comes from a rebuilt internal stack rather than a single trick. Datalab re-architected Marker 2 around three components: Surya OCR 2, a compact 20-million-parameter layout model, and a rewritten pdftext engine that runs three times faster than its predecessor. Together they let the tool sustain 2.9 pages per second on a single Nvidia B200 GPU while running in balanced mode.

Marker 2 also ships with tiered conversion paths rather than a one-size-fits-all pipeline. Balanced mode hits 76.0% on olmOCR-bench, fast mode drops to 66.6% for higher throughput, and a --disable_ocr option sacrifices accuracy (43.6%) for raw speed, hitting 23.7 pages per second when OCR isn't needed at all. That flexibility matters for teams processing large batches of born-digital files where full OCR is often unnecessary overhead.

How it stacks up against rivals

Against Docling, another open source contender, the gap is stark: Docling scores 50.3% on olmOCR-bench at 2.1 pages per second, well behind Marker 2's balanced-mode numbers on both accuracy and speed. Datalab is positioning Marker 2 as a tool that no longer forces users to trade one for the other.

The rewrite covers input formats beyond raw PDFs and scanned images, including PPTX, DOCX, XLSX, HTML, and EPUB, with output options in markdown, JSON, HTML, or chunked text suited for retrieval-augmented generation pipelines. That range positions Marker 2 less as a narrow OCR utility and more as a general-purpose document ingestion layer for downstream AI systems.

On the engineering side, Marker 2 requires Python 3.10 or later and has migrated its tooling from Poetry to uv, with hatchling now serving as the build backend. It also adds full CPU support and is device-aware out of the box, defaulting to balanced mode on GPU and fast mode on CPU or Apple's MPS backend, so the same codebase adapts automatically to whatever hardware it's running on.

For teams building document-heavy AI pipelines, whether feeding LLMs, populating vector databases, or automating enterprise data extraction, the combination of higher accuracy and multiples-faster throughput could meaningfully cut infrastructure costs at scale.