MinerU

MinerU is an open-source document parsing project from the OpenDataLab group. Version 4.0 combines document parsing, a local document library, and service tools in one workflow for document conversion, application integration, and agent reading. The latest listed release is MinerU 4.0.12.

What It Is

MinerU converts complex documents into structured, readable output. Supported inputs include PDF, images, DOC/DOCX, PPT/PPTX, XLS/XLSX, RTF, OpenDocument formats, EPUB, OFD, HTML/MHTML, and CSV/TSV. Its agent guide describes it as a command-line document reader for agents, and states that it is not a RAG framework, vector database, or chat-with-doc application.

Parsing Tiers and Runtimes

Four parsing tiers are documented: Flash for fast previews and indexing, Basic for OCR and model-based parsing, and Standard and Advanced for more demanding layouts and higher quality. PDFs and images support all four tiers, while Office, OpenDocument, EPUB, OFD, HTML/MHTML and CSV/TSV use local Flash native parsing. Small models run on ONNX or Torch, and the vision-language model runs on llama.cpp, vLLM, or LMDeploy. The default install uses ONNX CPU inference and llama.cpp in Vulkan mode, and a full extra is available for NVIDIA GPUs.

Document Library and Agent Reading

The local document library discovers files, caches results, searches indexed content, and lets readers continue by page or block. Results carry stable locators of the form doc/tier/page/block for citations and follow-up reads. A mineru agent skill can be installed so coding agents prefer MinerU for reading documents.

Interfaces and Outputs

MinerU offers a CLI, a Python SDK (DoclibClient), a V1 HTTP API, stateless batch conversion, a multi-service Router, and a Gradio-based WebUI. One document model renders to Markdown, HTML, LaTeX, DOCX, EPUB, PDF, Structured Content, and Content List V1/V2, though each CLI or API exposes only a subset of these exports.

Privacy Model

Documents are parsed locally by default. Remote parsing happens only when explicitly configured or requested with the --remote option. Anonymous telemetry can be inspected, enabled, or disabled by the user.