MinerU

MinerU is an open-source, AI-powered PDF content extrACTion toolkit designed for developers, researchers, and technical teams who need high-fIDElity, structured access to document content. Unlike generic PDF readers or simple OCR tools, MinerU treats document paRSIng as a multi-stage AI Workflow — combining layout analysis, multimodal reCognition, and semantic post-processing to preserve meaning, hierarchy, and fORMatting across diverse PDF types including scanned documents, LaTeX-generated papers, and complex technical manuals.

What it does

MinerU transforms unstructured PDFs into machine-ready, semantically organized ouTPUts. It goes beyond bASIc text extraction by detecting and separating visual elements like figures, tables, and equations — then converting them into APPropriate digital representations (e.g., LaTeX for formulas, HTML/Markdown for tables, alt-text descriptions for images). Its output supports multiple intermediate and final formats, including middle-json (a unified internal representation), Layout, Span, Content List, and human-editable markdown — making it ideal for integration into data pipelines, LLM pretraining Workflows, or knowledge base construction.

Key capabilities

• Intelligent layout detection: Uses LayoutLMv3 and custom fine-tuned models to identify titles, paRAGraphs, lists, captions, and regions with high spatial accuracy.
• Math-aware formula processing: Leverages a YOLOv8-based detector and UniMERNet recognizer to locate and convert both inline and display-mode equations into production-grade LaTeX.
• Multimodal handling: Processes Embedded images, vector graphics, and tabular data while preserving their logical relationship to surrounding text.
• Noise suppression: Automatically rEMOves headers, footers, page numbers, footnotes, and watermarks without manual intervention.
• Robust OCR & encoding recovery: Integrates PaddleOCR and custom heuristics to handle garbled text, font substitution artifacts, and low-resolution scans.
• quality assurance loop: Includes built-in evaluation using human-annotated test sets and visualization tools for error analysis and model refinement.

Who it is for

MinerU serves technical users: NLP engineers building domain-specific corpora, academic researchers curating literature datasets, legal or compliance teams extracting clause-level Information, and software teams automating documentation ingestion. It’s especially valuable when input documents contain mixed content — such as scientific papers with equations and figures, engineering schematics with annotations, or regulatory filings with nested tables.

How to get started

The project is hosted on GitHub and distributed under an open license. Users can run MinerU locally via Python CLI or integrate its modules directly into Python applications. Docker support simplifies deployment, and the modular architecture allows swapping components (e.g., swapping OCR backends or layout models). Documentation includes usage examples, configuration guides, and Sample pipelines for common use cases like arXiv paper ingestion or contract clause extraction.

Pricing and access caveat

MinerU is fully free and open-source with no usage liMITs or paywalls. However, local execution requires moderate GPU resources for optimal speed — particularly during layout and formula detection stages. cloud-hosted versions or managed APIs are not officially provided; users must self-host or adapt the code. Performance may vary depending on PDF complexity, scan quality, and hardware configuration.