Nougat (Neural Optical Understanding for Academic Documents) is a model from Meta AI that converts scanned or PDF academic papers directly into clean Markdown โ including properly formatted mathematical equations and tables. Published by Facebook Research, the open-source project has over 10,000 stars on GitHub and remains the reference approach for making scientific PDFs machine-readable.
Traditional OCR pipelines lose the structure of scientific documents: formulas become garbage characters and reading order falls apart. Nougat instead treats the page as an image and uses a Vision Transformer encoder-decoder to “read” it end to end, emitting Markdown with LaTeX math. It was trained on millions of pages from arXiv and PubMed Central, needs no external OCR engine or document parser, and ships as a simple command-line tool (also on PyPI as nougat-ocr) that processes single PDFs or batches of them, with a Hugging Face demo for quick tests. Output quality on formulas and multi-column layouts is the reason it is embedded in so many academic-PDF pipelines: downstream tools receive Markdown they can index, quote, and feed to language models directly.
Free. Nougat is open source under the MIT license (model weights under their research terms on Hugging Face). You run it on your own GPU with no fees or subscriptions.
Researchers, digital-library builders, and RAG engineers who need to ingest scientific literature โ turning equation-heavy PDFs into structured text that search engines, note-taking apps, and LLMs can actually use.