Marker

Marker turns PDFs, images, and office formats such as DOCX, PPTX, XLSX, EPUB, and HTML into Markdown, JSON, HTML, or retrieval-ready chunks, keeping tables, equations, code, links, and figures and dropping page headers and footers. It reads the embedded text layer where that is reliable and calls Datalab's Surya vision-language model for layout, OCR, inline math, and damaged pages, served by a local vLLM container on NVIDIA GPUs or by llama.cpp on CPUs and Apple Silicon. A balanced mode favors quality on GPUs, a fast mode keeps model calls to a minimum on CPUs, and an optional flag hands hard sections to an LLM such as Gemini, Claude, or Ollama to join tables split across pages and read forms. It ships single-file and batch CLIs, a Python API, and a small Streamlit app. Install it with pip install marker-pdf on Python 3.10 or later with PyTorch. Code is Apache 2.0; the weights ship under a modified OpenRAIL-M license that costs nothing for research, individuals, and startups below 5 million dollars of funding or revenue, while larger companies pay Datalab. It suits RAG builders and anyone digitizing papers or books.