mineru

MinerU is a high-accuracy document parsing engine designed to convert complex document formats into machine-readable structured data for downstream LLM, RAG, and agent workflows. It natively supports PDF, images, DOCX, PPTX, and XLSX inputs, and outputs Markdown or JSON with reading-order preservation. ## Key Capabilities - **Multi-format native parsing**: Handles PDF, scanned documents, images, DOCX, PPTX, and XLSX without requiring conversion to PDF first. - **Structure preservation**: Removes headers, footers, footnotes, and page numbers; outputs headings, paragraphs, lists, tables, and text in human-readable reading order for single-column, multi-column, and complex layouts. - **Formula and table extraction**: Automatically converts formulas to LaTeX and tables to HTML, including cross-page table merging and table content recognition. - **OCR engine**: Automatically detects scanned or garbled PDFs and enables OCR, supporting 109 languages. The pipeline backend uses PP-OCRv6 as of version 3.4. - **Dual inference backends**: Offers three modes — `pipeline` (rule-based, fast, CPU/GPU capable), `vlm-engine` (VLM-based, high accuracy, supports vLLM/LMDeploy/mlx), and `hybrid-engine` (combines native text extraction with VLM for low hallucination). - **Model choices**: VLM models include MinerU2.5-Pro series; the 3.3 release introduced an `effort` parameter (medium/high) to balance speed and accuracy. ## Integration and Deployment - **SDKs and interfaces**: Python/Go/TypeScript SDKs, CLI, REST API, Docker, and a Gradio WebUI. - **Frameworks**: Native integrations with LangChain, LlamaIndex, RAGFlow, Dify, FastGPT, Flowise, and MCP Server for AI coding tools like Cursor, Claude Desktop, and Windsurf. - **Routing and scaling**: Includes `mineru-api` for synchronous and asynchronous task endpoints, and `mineru-router` for unified entry and load balancing across multiple services and GPUs. Multi-threaded concurrent inference is supported. - **Platform support**: Runs on Windows, Linux, and macOS, including pure CPU environments and GPU/MPS acceleration. - **Private/offline deployment**: Models can be downloaded and cached locally; supports domestic AI chips such as Ascend, Cambricon, Enflame, MetaX, Moore Threads, Kunlunxin, Iluvatar, Hygon, Biren, and T-Head. ## License and Origins MinerU originated from the InternLM pretraining project and is maintained by OpenDataLab. Starting with version 3.1.0, it uses a custom open-source license based on Apache 2.0 (previously AGPLv3). An online web app, desktop client, and API are available on mineru.net, with demos on Hugging Face and ModelScope.