MinerU is a high-accuracy document parsing engine designed to transform complex documents into structured formats suitable for large language model workflows, retrieval-augmented generation systems, and agentic applications. The project converts PDFs, DOCX, PPTX, XLSX files, images, and web pages into markdown or JSON output with support for 109 languages through its dual VLM and OCR engine architecture.
The repository has grown to 73,389 stargazers and 6,160 forks as of the most recent tracking period, with consistent community engagement. GitGenius tracking shows a median issue and pull request response latency of 0.0 hours across 1,815 items, indicating rapid triage and response cycles. The most active contributor, myhloli, has logged 3,991 events, followed by dt-yy with 257 events and drunkpig with 129 events. Bug reports represent the dominant issue category with 1,129 tracked items, followed by enhancement requests at 376 items and general questions at 56 items. The project shares contributors with related repositories including PaddleOCR, Dify, and RAGFlow, indicating integration within a broader ecosystem of document processing and AI workflow tools.
MinerU's core parsing capabilities include native support for DOCX, PPTX, and XLSX formats with formula-to-LaTeX conversion and table-to-HTML transformation. The system handles scanned documents, handwriting recognition, multi-column layouts, and cross-page table merging while maintaining human reading order and automatically removing headers and footers. The dual engine approach combines vision language models with OCR for robust document understanding across diverse input types.
The platform offers multiple deployment and integration pathways. It provides MCP Server support for AI coding tools like Cursor and Claude Desktop, integrates with RAG frameworks including LangChain, LlamaIndex, RAGFlow, and Dify, and offers Python, Go, and TypeScript SDKs alongside CLI, REST API, and Docker deployment options. A web version at mineru.net provides zero-installation access, while desktop clients and Gradio WebUI interfaces serve different user needs. The system supports multiple inference backends including pipeline, vlm-engine, and hybrid-engine configurations, with compatibility for domestic AI chips including Ascend, Cambricon, Enflame, and others.
Recent releases demonstrate active development focused on performance optimization and capability expansion. Version 3.4 upgraded the OCR model to PP-OCRv6, achieving approximately 11 percent accuracy improvement on OmniDocBench v1.6 while doubling OCR processing speed. Version 3.3 introduced an effort parameter for the hybrid backend with medium and high parsing strength levels, delivering 35 to 220 percent speed improvements across different platforms while maintaining parsing accuracy. Version 3.1.0 transitioned the project from AGPLv3 to a custom MinerU Open Source License based on Apache 2.0, reducing adoption friction for both community and commercial deployments. The VLM model was upgraded to MinerU2.5-Pro-2604-1.2B with support for image and chart parsing, truncated paragraph merging, and cross-page table merging capabilities.
The project is classified across multiple domains including data mining, machine learning, AI platforms, data science, deep learning, workflow management, model training, data processing, and MLOps, reflecting its position as a comprehensive document understanding solution for enterprise and research applications.
