Marker 2 Open-Source Document Parsing: How Datalab's New Pipeline Outperforms MinerU, Docling and LiteParse

Marker 2 Sets a New Bar for Open-Source Document Parsing

Datalab has released Marker 2, a ground-up rewrite of its popular open-source document parsing library, and the performance numbers are turning heads across developer and data engineering communities. In benchmark testing on olmOCR-bench — one of the most rigorous public evaluations for optical character recognition and document understanding — Marker 2 achieved a score of 76.0, outpacing established competitors including MinerU, Docling, and LiteParse on both accuracy and throughput. For developers, privacy professionals, and IT decision-makers evaluating open-source document parsing tools for GDPR-compliant, on-premise pipelines, this release changes the conversation significantly.

The headline figure is throughput: running on a single B200 GPU, Marker 2 sustains 2.9 pages per second — a rate the Datalab team reports as more than five times faster than MinerU's pipeline backend. That kind of performance gap matters enormously in enterprise and research settings where millions of pages need to be processed without sending sensitive documents to third-party cloud APIs. As data sovereignty regulations tighten across Europe and beyond, the ability to run high-accuracy document extraction entirely on local infrastructure is increasingly a compliance requirement, not just a preference.

How the Three-Mode Pipeline Architecture Works

Marker 2's most significant architectural innovation is its three-mode pipeline design. Rather than applying the same processing chain to every document regardless of complexity, Marker 2 intelligently routes documents through one of three modes: a lightweight path for clean, machine-readable PDFs; an intermediate path for documents requiring layout analysis; and a full OCR path for scanned or image-heavy content. This tiered approach means the system doesn't waste GPU cycles running heavy neural inference on a clean text-layer PDF, while simultaneously ensuring that complex scanned documents receive the full treatment they require.

This kind of adaptive routing is a meaningful step forward compared to tools like Docling — IBM's open-source document understanding library — which applies a more uniform processing approach. According to Datalab's own benchmarking, Marker 2 beats Docling on both accuracy metrics and processing speed, which is notable given that Docling has attracted significant enterprise adoption since its open-source release, particularly in European organisations building GDPR-compliant document workflows. IBM's Docling project, tracked on GitHub, has accumulated substantial community traction — making Marker 2's competitive claims all the more striking.

The adaptive routing also has practical implications for cost efficiency. In cloud-hosted environments, compute cost correlates directly with inference time. A tool that processes five times more pages per GPU-hour translates to dramatically lower operational costs — a factor that IT decision-makers evaluating total cost of ownership will want to model carefully.

olmOCR-Bench Results: Where Each Tool Stands

The olmOCR-bench benchmark, developed to evaluate document understanding across a range of real-world PDF types — including academic papers, financial reports, legal documents, and scanned images — has become a de facto standard for comparing PDF extraction and OCR tools. Marker 2's score of 76.0 positions it clearly ahead of the pack in this evaluation, though it is important to note that benchmark scores always reflect a particular distribution of test documents, and real-world performance can vary by document type and domain.

For context, the open-source OCR and document parsing landscape has grown rapidly. MinerU, developed and maintained by Shanghai AI Laboratory, has been a popular choice for research pipelines due to its strong layout analysis capabilities. LiteParse takes a different approach, optimising for lightweight deployments rather than maximum accuracy. Docling, backed by IBM Research, has positioned itself as an enterprise-grade option with strong support for structured document formats.

Why Open-Source Document Parsing Matters for GDPR Compliance and Data Sovereignty

For organisations operating under GDPR — or any of the growing number of similar data protection frameworks — the ability to process documents without relying on commercial cloud APIs is increasingly strategic. Every page sent to a third-party extraction API is a potential data transfer, with all the compliance obligations that entails: data processing agreements, transfer impact assessments, and scrutiny over where data is stored and processed. According to guidance published by the European Data Protection Board, organisations must have clear legal bases and appropriate safeguards for any personal data transferred outside their control, even for processing purposes.

This is precisely the context in which tools like Marker 2 become significant beyond their benchmark scores. When a law firm needs to extract text from thousands of client documents, or a healthcare provider wants to digitise patient records, or a financial institution is parsing regulatory filings, sending those documents to an external API may be impractical or legally problematic. Running Marker 2 on local GPU infrastructure — or within a private cloud environment — eliminates that exposure entirely.

Research published through arXiv on document AI pipelines has consistently highlighted that the gap between open-source and proprietary solutions has narrowed substantially over the past two years, driven largely by improvements in vision-language models and efficient inference frameworks. Marker 2's performance on olmOCR-bench suggests that open-source document parsing has now crossed a threshold where it is competitive with commercial alternatives on raw quality metrics, not merely on cost or control grounds.

Which Open-Source Document Parsing Tool Fits Your Use Case?

Choosing between Marker 2, MinerU, Docling, and LiteParse is not simply a matter of picking the highest benchmark score. Each tool has been designed with different constraints and priorities in mind, and the right choice depends on the specifics of your pipeline.

Marker 2 is the strongest general-purpose choice for teams that need high accuracy across a mixed document corpus and have access to GPU hardware — either on-premise or in a private cloud. Its three-mode pipeline means it handles diverse document types efficiently without manual routing logic. For organisations processing large volumes of varied documents under data sovereignty constraints, it is currently the most compelling open-source option.

MinerU remains a solid choice for research-oriented pipelines, particularly those focused on academic and scientific documents. Its layout analysis capabilities have been refined specifically for structured academic content, and the existing community and tooling around it are substantial. However, its throughput disadvantage relative to Marker 2 will be a real cost factor at scale.

Docling is worth considering for enterprise teams already operating within IBM's ecosystem or those that prioritise a well-documented, modular architecture with strong structured document support. Its backing by IBM Research means it is likely to receive continued investment. However, both its accuracy score and reported throughput now lag behind Marker 2 according to Datalab's benchmarking, as covered in MarkTechPost's analysis.

LiteParse occupies a different niche: edge deployments, resource-constrained environments, or use cases where document volume is moderate and accuracy requirements are less demanding. It is not competing with Marker 2 on performance — but that

Originally reported by MarkTechPost. Summarised and curated by European Purpose.