AI News AI资讯 3h ago Updated 2h ago 更新于 2小时前 47

Datalab Marker v2 vs MinerU, Docling, and Liteparse: Benchmark Breakdown Datalab Marker v2 与 MinerU、Docling 和 Liteparse 对比:基准测试详解

Datalab released Marker v2, a complete rewrite of its open-source document conversion pipeline, integrating Surya OCR 2, a 20M-parameter layout model, and a rebuilt pdftext engine. In balanced mode on a single B200 GPU, Marker v2 achieves 76.0% accuracy on the olmOCR-bench with a throughput of 2.9 pages per second, significantly outperforming MinerU (0.54 pg/s) and Docling (2.1 pg/s). The architecture introduces three distinct modes: balanced (highest quality via VLM), fast (cost-effective via l Datalab发布Marker v2,基于Surya OCR 2、20M参数布局模型和加速3倍的pdftext重构,支持多格式文档转Markdown/JSON。 在olmOCR-bench基准测试中,Marker 2平衡模式得分76.0%,在B200 GPU上吞吐量达2.9页/秒,显著优于MinerU和Docling。 提供三种转换路径:平衡模式(高精度)、快速模式(低成本CPU友好)及禁用OCR模式(纯文本提取),默认根据设备自动选择。 架构采用多线程CPU共享单一Surya推理服务器,实现吞吐量随服务器容量扩展而非受限于单进程显存,解决并发瓶颈。 相比全页VLM方案,Marker 2在保持

65
Hot 热度
70
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Datalab released Marker v2, a complete rewrite of its open-source document conversion pipeline, integrating Surya OCR 2, a 20M-parameter layout model, and a rebuilt pdftext engine.
  • In balanced mode on a single B200 GPU, Marker v2 achieves 76.0% accuracy on the olmOCR-bench with a throughput of 2.9 pages per second, significantly outperforming MinerU (0.54 pg/s) and Docling (2.1 pg/s).
  • The architecture introduces three distinct modes: balanced (highest quality via VLM), fast (cost-effective via lightweight detectors), and disable_ocr (pure CPU text extraction).
  • Marker v2 demonstrates superior scalability by decoupling CPU workers from a shared Surya inference server, allowing throughput to scale with server capacity rather than per-process VRAM.
  • While full-page VLMs like Chandra 2 still lead in raw accuracy for complex scans, Marker v2’s balanced mode rivals API-based solutions like Gemini Flash 3.5 for born-digital PDFs at a fraction of the latency and cost.

Why It Matters

This release establishes a new benchmark for high-throughput, high-fidelity document parsing, offering AI practitioners a viable alternative to expensive or slow proprietary APIs for processing large volumes of PDFs. The ability to achieve near-VLM accuracy with significantly higher throughput on consumer-grade hardware (or single GPUs) lowers the barrier for enterprise RAG systems that require precise structural understanding of documents. Furthermore, the clear trade-offs between accuracy and speed across different modes allow teams to optimize their pipelines based on specific use cases, such as batch processing versus real-time ingestion.

Technical Details

  • Core Architecture: Marker v2 is built around three key components: Surya OCR 2 for optical character recognition, a 20M-parameter layout detection model for identifying columns, tables, and headers, and an optimized pdftext engine that is 3x faster than its predecessor.
  • Conversion Modes: The tool offers three paths: "balanced" uses the Surya VLM for layout and re-OCR when embedded text is poor (76.0% score); "fast" uses a lightweight rf-detr/onnx detector with minimal VLM usage (66.6% score); and "disable_ocr" performs pure text-layer extraction on CPU (43.6% score, 23.7 pg/s).
  • Benchmark Performance: Evaluated on Allen AI’s olmOCR-bench (1,403 PDFs, ~8,400 tests), Marker v2 balanced mode scored 76.0% overall and 83.5% on born-digital PDFs. It sustained 2.9 pages per second on a B200 GPU, compared to MinerU’s 0.54 pg/s and Docling’s 2.1 pg/s.
  • Scalability Design: The system utilizes many thin CPU workers sharing a single Surya inference server. This architecture budgets VLM concurrency across workers, enabling throughput to scale with server capacity rather than being limited by per-process VRAM constraints.
  • Compatibility & Changes: Requires Python 3.10+ and has shifted packaging from Poetry to uv/hatchling. Structured-extraction converters were removed, directing users to hosted APIs or LLM workflows for those features.

Industry Insight

  • Hybrid Pipeline Adoption: Organizations should consider adopting hybrid strategies where "fast" or "disable_ocr" modes handle high-volume, simple documents, while "balanced" mode is reserved for complex, scanned, or math-heavy documents requiring high fidelity.
  • Cost-Efficiency in RAG: The significant throughput advantage over competitors like MinerU suggests that self-hosted Marker v2 can drastically reduce infrastructure costs for RAG systems processing large document corpora, especially when leveraging CPU-only modes for less critical data.
  • Benchmark Contextualization: While Marker v2 shows strong results, teams must validate these findings against their own specific document mixes, as performance varies significantly between born-digital PDFs and scanned images. The gap between pipeline-based tools and full-page VLMs remains relevant for edge cases involving complex layouts or poor-quality scans.

TL;DR

  • Datalab发布Marker v2,基于Surya OCR 2、20M参数布局模型和加速3倍的pdftext重构,支持多格式文档转Markdown/JSON。
  • 在olmOCR-bench基准测试中,Marker 2平衡模式得分76.0%,在B200 GPU上吞吐量达2.9页/秒,显著优于MinerU和Docling。
  • 提供三种转换路径:平衡模式(高精度)、快速模式(低成本CPU友好)及禁用OCR模式(纯文本提取),默认根据设备自动选择。
  • 架构采用多线程CPU共享单一Surya推理服务器,实现吞吐量随服务器容量扩展而非受限于单进程显存,解决并发瓶颈。
  • 相比全页VLM方案,Marker 2在保持接近顶级API精度的同时,避免了逐页调用成本,为大规模文档处理提供了高性价比替代方案。

为什么值得看

本文详细对比了开源文档解析工具的最新性能数据,揭示了在精度与吞吐量之间取得平衡的技术路径,对构建RAG或数据清洗流水线有直接参考价值。它展示了如何通过架构优化(如共享推理服务)突破传统单进程处理的性能瓶颈,为AI从业者选型提供了量化依据。

技术解析

  • 核心组件升级:Marker v2由Surya OCR 2(20M参数快速布局模型)和重写的pdftext组成,后者速度提升3倍。整体架构围绕这三个经过数月迭代的核心组件构建,支持PDF、图片、PPTX等多种格式输入。
  • 基准测试表现:基于Allen AI的olmOCR-bench(1,403份PDF,涵盖数学公式、表格结构等8类测试),Marker 2平衡模式总体得分76.0%,原生数字PDF得分83.5%。在B200 GPU上,其吞吐量达到2.9页/秒,是MinerU后端(0.54页/秒)的5倍以上,且精度更高。
  • 多模式策略与硬件适配:提供“balanced”(GPU优先,最高质量)、“fast”(轻量级检测器+最小化VLM使用,成本低)和“–disable_ocr”(纯CPU提取,23.7页/秒但精度较低)。系统默认根据设备类型自动选择模式,并支持完全在CPU上运行20M布局模型以恢复列、表结构。
  • 并发架构创新:通过让多个轻量级CPU工作线程共享单个Surya推理服务器,父进程协调VLM并发预算。这种设计使吞吐量取决于服务器容量而非单进程VRAM,将平衡模式的实际速率从单流的0.3页/秒提升至2.9页/秒。
  • 工程变更:要求Python 3.10+,打包工具从Poetry移至uv(hatchling后端),移除了结构化提取转换器,建议用户使用托管API或LLM工作流替代。

行业启示

  • 文档解析不再是简单的OCR任务:随着RAG应用的普及,文档的结构化还原(如表格、公式、阅读顺序)成为关键痛点。Marker 2证明通过专门的布局模型和混合策略(文本层+选择性OCR)可以在保证精度的同时大幅提升效率,优于纯端到端VLM方案。
  • 吞吐量与精度的权衡需精细化设计:企业应根据文档类型(原生数字vs扫描件)和业务需求(实时性vs准确性)选择不同模式。对于大规模非关键文档,CPU友好的快速模式或禁用OCR模式可提供极高的性价比;而对于高精度需求,GPU加速的平衡模式提供了接近API服务的体验。
  • 开源生态的竞争焦点转向架构优化:在基础模型能力趋同的背景下,系统级的架构优化(如共享推理服务、设备感知调度)成为拉开性能差距的关键。开发者应关注此类工程实践,而非仅依赖单一模型的参数规模。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Benchmark 基准测试 Evaluation 评测