MinerU

OpenDataLab's MinerU converts documents into text that language models can use, rebuilding reading order and keeping tables, formulas and figures intact. Version 4.0 sorts parsing into four quality tiers: a fast native pass for previews and indexing, a basic tier with OCR and layout models, and standard and advanced tiers that call a vision-language model for hard pages. PDFs and images can use every tier, while Word, PowerPoint, Excel, OpenDocument, EPUB, OFD, HTML and CSV files go through the fast local parser. Results can be saved as Markdown, JSON content lists, LaTeX, HTML, DOCX and more. It ships a CLI, a Python SDK, an HTTP API, a multi-server router and a Gradio web interface, and stays local unless you opt into remote parsing. The default install runs ONNX on the CPU and llama.cpp over Vulkan, needing 2 GB of RAM for basic and 8 GB for VLM tiers; NVIDIA owners can add vLLM or LMDeploy. Licensing is Apache 2.0 with extra terms: commercial use is allowed, online services must credit MinerU, and firms above 100 million monthly users or 20 million dollars in monthly revenue need a separate license. Suits RAG ingestion.