Marker 2 Document Conversion Pipeline Achieves 5x Throughput Boost over MinerU
Datalab released Marker 2, a complete rewrite of its open-source document parsing pipeline using Surya OCR 2 and a 20M-parameter layout model. Benchmark tests on olmOCR-bench show Marker 2 reaching 76.0% overall accuracy at 2.9 pages per second on a single B200 GPU.
Impact: Medium
Why it matters
Engineers building RAG pipelines can process document datasets significantly faster on lower hardware budgets without trading off structural precision.
TL;DR
- 01Marker 2 achieves 76% accuracy on olmOCR-bench while running 5.4x faster than MinerU
- 02Decoupled architecture lets lightweight CPU workers share a single Surya VLM inference server
Key facts
- Balanced Mode Speed (self-reported)
- 2.9 pages/sec (B200 GPU)
- Balanced Mode Accuracy
- 76.0% overall (olmOCR-bench)
- CPU-Only Speed
- 23.7 pages/sec (--disable_ocr)
- Licensing Cap
- Free weights under $5M funding/revenue
Architectural Overhaul
Datalab rebuilt Marker 2 around three updated foundation layers: Surya OCR 2, a fast 20M-parameter layout detection model, and an upgraded pdftext engine that runs 3x faster than previous iterations. A key optimization is the decoupled inference model where thin CPU worker processes offload vision tasks to a centralized Surya server.
Benchmark Results on olmOCR-bench
On Allen AI's olmOCR-bench (1,403 PDFs and ~8,400 pass/fail tests), Marker 2 balanced mode achieved 76.0% overall score at 2.9 pages per second on a single NVIDIA B200 GPU. This outperforms MinerU pipeline backend (72.7% at 0.54 pg/s) and Docling (50.3% at 2.1 pg/s).
Operational Modes and Licensing
balanced: GPU-optimized layout parsing with selective OCR (76.0% accuracy).fast: CPU/MPS mode with minimal VLM overhead (66.6% accuracy, 7.4 pg/s).--disable_ocr: Pure CPU text layer extraction (43.6% accuracy, 23.7 pg/s).
Licensing: Code is Apache 2.0; model weights use OpenRAIL-M (free for startups under $5M revenue/funding).
Try it in 2 minutes
pip install marker-pdf
marker_single sample.pdf --mode balancedbash
✓ When to use
- Converting large PDF or document sets into clean Markdown for RAG ingestion
- Extracting structured text from digital and scanned PDFs on local GPU clusters
✕ When NOT to use
- Ultra-high throughput single-pass CPU text extraction without structure (use LiteParse instead)
- Commercial deployment with over $5M in revenue without a commercial model weight license
What to do today
- Test Marker 2 balanced mode against existing PDF extractors on your local dataset
- Verify commercial licensing requirements if company revenue exceeds $5M
Sources