Datalab’s Marker 2 Revolutionizes Document Conversion: Up to 5x Faster and More Accurate Than Competitors

A New Era for Document Conversion: Introducing Datalab’s Marker 2

For readers tracking the shift, Datalab has unveiled Marker 2, a complete overhaul of its open-source document conversion pipeline. This next-generation tool promises a significant leap in efficiency and accuracy, capable of transforming a wide array of file types—including PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB—into structured formats like markdown, JSON, HTML, or discrete chunks. Built upon the foundation of Surya OCR 2, a rapid 20M-parameter layout model, and a re-engineered pdftext component that boasts 3x faster performance, Marker 2 is set to redefine expectations in document processing.

Unpacking the Performance: olmOCR-bench Results

Meanwhile, The true power of Marker 2 is showcased through the rigorous olmOCR-bench, a third-party benchmark developed by Allen AI. Marker 2’s ‘balanced’ mode achieved an impressive 76.0% overall score, with an even higher 83.5% on born-digital PDFs. Crucially, it sustains an remarkable throughput of 2.9 pages per second on a single B200 GPU. This performance metric is particularly striking when compared to its rivals:

• MinerU’s pipeline backend: Scores 72.7% at just 0.54 pages per second. Marker 2 delivers over 5 times MinerU’s throughput at a superior accuracy.
• Docling: Achieves 50.3% at 2.1 pages per second on the same benchmark harness.

Marker 2’s Flexible Conversion Paths

Marker 2 introduces three distinct conversion modes, allowing users to tailor performance to their specific needs:

• Balanced Mode: This is the flagship mode, where the Surya VLM handles layout, and the entire page is re-OCR‘d when embedded text quality is low. It offers the highest quality, performs best on GPU, and scored 76.0% on olmOCR-bench.
• Fast Mode: Utilizing a lightweight rf-detr/onnx layout detector and pdftext with minimal VLM intervention, this mode is significantly cheaper to run. It achieved 66.6% on the benchmark.
• –disable_ocr Mode: For scenarios requiring pure text-layer extraction without any VLM calls, this mode runs entirely on CPU. While scoring 43.6%, it boasts an impressive 23.7 pages per second throughput.

In practical terms, Beyond these modes, Marker 2 features default device awareness (balanced on GPU, fast on CPU/MPS) and robust full CPU support. A key architectural improvement involves many thin CPU workers sharing a single Surya inference server, allowing throughput to scale with server capacity rather than per-process VRAM.

Important Upgrade Notes

For those considering an upgrade, it’s worth noting the following breaking changes:

• Python 3.10+ is now a requirement.
• Packaging has transitioned from Poetry to uv, with hatchling as the build backend.
• The structured-extraction converter and extractors have been removed, with Datalab directing users to their hosted API or a –use_llm workflow instead.

Head-to-Head: Marker 2 Against the Competition

For example, The olmOCR-bench, a comprehensive benchmark of 1,403 PDFs with roughly 8,400 unit tests, provides a clear picture of Marker 2’s standing. While these figures are from Datalab’s own reproducible runs, they offer valuable insights.

Marker 2 vs. MinerU

Architecturally similar, both Marker 2 and MinerU selectively OCR based on the PDF text layer. Marker 2’s balanced mode leads MinerU with 76.0% vs. 72.7% overall. On born-digital documents, they are closely matched (83.5% vs.

83.3%). The critical differentiator is throughput: Marker 2’s balanced mode is 5.4 times faster (2.9 pg/s vs. 0.54 pg/s) at a higher score. Marker’s fast mode pushes this to 13.7 times faster, though with a slightly lower score.

Marker 2 vs. Docling

That said, Marker 2 demonstrates a significant lead over Docling in GPU pipelines. Marker 2’s balanced mode outshines Docling with 76.0% vs. 50.3% overall, and 83.5% vs.

64.0% on born-digital documents. It also processes faster at 2.9 pg/s compared to Docling’s 2.1 pg/s. However, Docling offers strengths in governance (MIT-licensed, IBM Research origin, LF AI & Data Foundation project) and broader format support, extending to audio and email.

Marker 2 vs. LiteParse

LiteParse, a Rust document parser from the LlamaIndex team, operates on a different axis. On CPU, LiteParse scores 22.4% overall (20.4% with OCR off) against Marker’s CPU-only 43.6%.

LiteParse’s strength is pure speed, reporting 1721 pg/s with OCR disabled—approximately 73 times faster than Marker’s CPU mode. However, this comes at the cost of structural understanding, as LiteParse lacks a layout model and struggles with non-linear documents, whereas Marker’s CPU mode still recovers structure.

Marker 2 vs. Full-Page VLM Tier

Interestingly, Datalab emphasizes that Marker is a pipeline, distinct from full-page Vision-Language Models (VLMs). While dedicated VLMs like Datalab’s Chandra 2 (85.8%) or Gemini Flash 3.5 (76.4%) generally offer superior accuracy for complex scans and math-heavy pages, Marker’s balanced mode significantly narrows the gap.

It’s only 0.4 points behind Gemini Flash 3.5 overall and even surpasses it on born-digital documents (83.5% vs. 79.1%), all without requiring per-page API calls.

Understanding Per-Category Performance

The chosen mode in Marker 2 directly impacts its failure profile across different document categories. For instance, ‘fast’ mode, which reads equations from the PDF text layer instead of VLM-OCRing them, sees arXiv math scores drop sharply from 83.9% to 23.4%.

The –disable_ocr mode scores 0.0% in math categories by design. Old scans remain the weakest category across all modes, peaking at 43.2%.

Crucial Consideration: Licensing

For commercial teams, licensing models present a key differentiator:

• Marker: The code is Apache 2.0. However, its model weights utilize a modified AI Pubs OpenRAIL-M license, which is free for research, personal use, and startups with under $5M funding/revenue. Commercial use beyond these thresholds requires a paid license.
• MinerU: Operates under the MinerU Open Source License (based on Apache 2.0 with additional conditions). A separate commercial license is necessary for usage exceeding 100M monthly active users (MAU) or $20M in monthly revenue. Online services built on MinerU must also disclose this fact.
• Docling: Is MIT-licensed, with model licenses tracked separately based on their original packages.
• LiteParse: Is open source from run-llama, with LlamaParse offered as a paid cloud service for more challenging documents.

Choosing the Right Tool for Your Use Case

Ultimately, the ideal document conversion tool isn’t solely determined by benchmark scores. Factors such as your document corpus type, available hardware, specific licensing requirements, and desired output format play crucial roles. Evaluating these systems against your own document sets is the most reliable way to determine which solution best fits your operational needs.

Expert Perspective

A practical read on Datalab Marker 2 starts with marker. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make Datalab Marker 2 a meaningful reference point across mode.

For decision-makers, the useful lens is not the headline alone but how balanced changes priorities once organizations have to respond.

Frequently Asked Questions

Why is Datalab Marker 2 important?

A New Era for Document Conversion: Introducing Datalab’s Marker 2For readers tracking the shift, Datalab has unveiled Marker 2, a complete overhaul of its open-source document conversion pipeline.

What impact could Datalab Marker 2 have?

This next-generation tool promises a significant leap in efficiency and accuracy, capable of transforming a wide array of file types—including PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB—into structured formats like markdown, JSON, HTML, or discrete chunks.

What should readers watch next with Datalab Marker 2?

Built upon the foundation of Surya OCR 2, a rapid 20M-parameter layout model, and a re-engineered pdftext component that boasts 3x faster performance, Marker 2 is set to redefine expectations in document processing.Unpacking the Performance: olmOCR-bench ResultsMeanwhile, The true power of Marker 2 is showcased through the rigorous olmOCR-bench, a third-party benchmark developed by Allen AI.

How does this relate to marker?

It connects because the article frames marker as one of the clearest areas where the topic may be felt in practice.

Source: https://www.marktechpost.com/2026/07/24/datalabs-marker-2-vs-mineru-docling-and-liteparse-76-0-on-olmocr-bench-at-5x-minerus-throughput/