Datalab Marker 2: Unpacking the Next-Gen Document Conversion Pipeline

Revolutionizing Document Conversion with Datalab Marker 2

At a glance, Datalab has officially launched Marker 2, a comprehensive rewrite of its popular open-source document conversion pipeline. This new iteration promises significant advancements in accuracy, speed, and flexibility, setting a new standard for transforming diverse file types into structured data formats. Marker 2 converts PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB files into markdown, JSON, HTML, or various chunks, making it an indispensable tool for data extraction and processing.

Meanwhile, Built upon a foundation of cutting-edge components, Marker 2 integrates Surya OCR 2, a powerful 20M-parameter fast layout model, and a re-engineered pdftext module that boasts a 3x speed improvement over its predecessor. These core enhancements contribute to its impressive performance, which has been rigorously tested against leading competitors.

Understanding Marker 2’s Core Innovations

Marker 2 introduces several key features and architectural changes designed to optimize document conversion across various use cases and hardware configurations.

Flexible Conversion Paths

In practical terms, Unlike its single-path predecessor, Marker 2 offers three distinct conversion modes, allowing users to balance quality, speed, and resource consumption:

• Balanced Mode: This mode leverages the Surya VLM for layout processing and re-OCRs entire pages when embedded text quality is poor. It delivers the highest quality, particularly effective on GPU, achieving a 76.0% overall score on olmOCR-bench.
• Fast Mode: Employing a lightweight rf-detr/onnx layout detector alongside pdftext, this mode uses minimal, surgical VLM intervention. It scores 66.6% on olmOCR-bench but offers a significantly cheaper and faster alternative.
• Disable OCR Mode: For scenarios prioritizing pure text-layer extraction without any VLM calls, this mode runs entirely on CPU. While scoring 43.6% on olmOCR-bench, it achieves an impressive 23.7 pages per second.

Enhanced Device Awareness and CPU Support

Marker 2 is now device-aware by default, intelligently selecting the balanced mode for GPUs and fast mode for CPUs/MPS. Crucially, it offers robust full CPU support, with the fast and disable_ocr modes capable of running without a GPU or an inference server. Even on CPU, the 20M layout model efficiently processes columns, tables, and headers.

Scalable Architecture for Superior Throughput

For example, A significant architectural overhaul contributes to Marker 2’s impressive throughput. It now employs multiple thin CPU workers that share a single Surya inference server.

The parent process intelligently manages VLM concurrency across these workers, allowing throughput to scale with server capacity rather than being limited by per-process VRAM. This innovation enables the balanced mode to sustain approximately 2.9 pages per second, a substantial leap from the ~0.3 pages per second single-stream rate on the same hardware.

Important Upgrade Considerations

Users planning an upgrade should note a few breaking changes: Python 3.10+ is now a requirement. The packaging system has transitioned from Poetry to uv, with hatchling as the build backend, though pip install marker-pdf remains unchanged. Additionally, the structured-extraction converter and extractors have been removed, with Datalab directing users to their hosted API or a –use_llm workflow for similar functionalities.

The olmOCR-bench Benchmark: A Third-Party Validation

That said, The core of Marker 2’s performance claims rests on the olmOCR-bench, a robust third-party benchmark developed by Allen AI. This comprehensive test suite includes 1,403 PDFs and approximately 8,400 pass/fail unit tests, covering critical aspects like math rendering, table structure, reading order, headers, footers, and older scanned documents. The overall score represents a macro-average across eight categories, computed using the official olmOCR-bench checker.

Notably while olmOCR-bench is from Ai2, all scores and throughput figures presented by Datalab are derived from their own reproducible runs using an open harness available in the Marker repository. This harness also includes competitor runners for MinerU, Docling, and LiteParse, ensuring a transparent comparison. However, Datalab emphasizes that results can vary based on document mix and hardware, recommending teams run the harness against their own specific corpus.

Performance Showdown: Marker 2 vs. the Competition

Marker 2’s performance is best understood in comparison to its contemporaries.

Marker 2 vs. MinerU

MinerU’s pipeline backend shares a similar architectural approach, reading the PDF text layer and selectively applying OCR. In terms of overall accuracy, Marker 2’s balanced mode leads with 76.0% against MinerU’s 72.7%. On born-digital documents, both are closely matched at 83.5% and 83.3% respectively.

However, The most significant divergence lies in throughput: Marker 2’s balanced mode sustains 2.9 pages per second, a remarkable 5.4x increase over MinerU’s 0.54 pages per second, all while achieving a higher accuracy score. Marker 2’s fast mode pushes this even further, sustaining 7.4 pages per second (roughly 13.7x MinerU’s rate), albeit with a slightly lower accuracy score.

Marker 2 vs. Docling

Docling, originating from IBM Research and hosted by the LF AI & Data Foundation, shows the widest performance gap among the GPU pipelines. Marker 2’s balanced mode significantly outperforms Docling, leading 76.0% to 50.3% overall and 83.5% to 64.0% on born-digital documents. Marker 2 also boasts faster processing at 2.9 pages per second compared to Docling’s 2.1 pages per second.

Meanwhile, While Marker 2 excels in accuracy and speed, Docling differentiates itself through its MIT license, robust governance, and broader format breadth, extending beyond documents to audio and email formats.

Marker 2 vs. LiteParse

LiteParse, a Rust document parser from the LlamaIndex team, operates on a different axis, prioritizing speed over deep structural recovery. On CPU, Marker 2’s CPU-only mode scores 43.6% compared to LiteParse’s 22.4% overall (20.4% with OCR off). This difference is due to Marker’s 20M layout model, which recovers structure even on CPU.

However, LiteParse with OCR disabled achieves an astonishing 1721 pages per second, roughly 73x Marker 2’s CPU mode. This highlights a clear trade-off: LiteParse offers extreme speed for plain text extraction but struggles with non-linear layouts due to the absence of a layout model.

Marker 2 vs. the Full-Page VLM Tier

Datalab emphasizes that Marker 2 is designed as a pipeline, distinct from full-page Visual Language Models (VLMs). While VLMs like Datalab’s hosted Chandra 2 (85.8%), Ai2’s olmOCR 2 (82.4%), and Gemini Flash 3.5 (76.4%) generally offer superior accuracy for highly complex documents, scans, and math-heavy pages, Marker 2’s balanced mode significantly narrows the gap.

For example, Remarkably, Marker 2’s balanced mode is only 0.4 points behind Gemini Flash 3.5 overall and even outperforms it on born-digital documents (83.5% vs. 79.1%), all without requiring a per-page API call. This positions Marker 2 as a highly competitive and efficient solution for many high-accuracy scenarios.

Navigating Licensing for Commercial Use

For commercial teams, licensing can be a critical differentiator among these systems:

• Marker: The code is Apache 2.0. However, model weights use a modified AI Pubs OpenRAIL-M license. This means commercial use of the weights requires a paid license for startups exceeding $5M in funding or revenue, while research, personal use, and smaller startups are free.
• MinerU: Operates under the MinerU Open Source License (based on Apache 2.0 with added conditions). A separate commercial license is necessary for usage above 100M monthly active users (MAU) or $20M in monthly revenue. Online services built on MinerU must also disclose this fact.
• Docling: Licensed under MIT, making it highly permissive. Model licenses are tracked separately in their original packages.
• LiteParse: An open-source project from run-llama. Its commercial counterpart for complex documents is LlamaParse, offered as a paid cloud service.

Key Takeaways for Document Processing Professionals

That said, Datalab Marker 2 represents a significant leap forward in open-source document conversion. Its blend of accuracy, speed, and architectural innovation makes it a strong contender in the field. Here are the core insights:

• Marker 2’s balanced mode achieves 76.0% on olmOCR-bench at 2.9 pages per second, outperforming MinerU’s pipeline throughput by over 5x.
• It decisively beats Docling in both accuracy (76.0% vs. 50.3%) and speed (2.9 pg/s vs. 2.1 pg/s).
• LiteParse excels in raw speed for pure text extraction (1721 pg/s with OCR off) but sacrifices structural recovery, scoring significantly lower than Marker’s CPU mode (20.4% vs. 43.6%).
• Marker 2’s fast mode with –disable_ocr offers a fully CPU-based solution at 23.7 pages per second, requiring no inference server.
• Licensing models vary widely: Docling is MIT, MinerU offers a free tier up to $20M monthly revenue, while Marker’s weights require a paid license for commercial entities above $5M in funding/revenue.
• All benchmark figures are reproducible via an open harness in the Marker repository, encouraging users to test against their own document sets.

Ultimately, the choice of tool depends on specific requirements: document types, available hardware, budget, licensing constraints, and desired output formats. Marker 2 provides a compelling, high-performance option that challenges even dedicated VLM solutions for many critical document processing tasks.

Expert Perspective

From an industry angle, the clearest signal around Datalab Marker 2 is how it may influence marker. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Datalab Marker 2 room to reshape expectations across mode over the near term.

For readers focused on practical impact, the best next step is to watch what changes around olmocr once attention turns into execution.

Frequently Asked Questions

Why does Datalab Marker 2 matter right now?

Revolutionizing Document Conversion with Datalab Marker 2At a glance, Datalab has officially launched Marker 2, a comprehensive rewrite of its popular open-source document conversion pipeline.

What broader change could Datalab Marker 2 signal?

This new iteration promises significant advancements in accuracy, speed, and flexibility, setting a new standard for transforming diverse file types into structured data formats.

What should the market watch next around Datalab Marker 2?

Marker 2 converts PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB files into markdown, JSON, HTML, or various chunks, making it an indispensable tool for data extraction and processing.Meanwhile, Built upon a foundation of cutting-edge components, Marker 2 integrates Surya OCR 2, a powerful 20M-parameter fast layout model, and a re-engineered pdftext module that boasts a 3x speed improvement over its predecessor.

Source: https://www.marktechpost.com/2026/07/24/datalab-marker-v2-vs-mineru-docling-and-liteparse-benchmark-breakdown/