Nougat

Nougat is a local AI PDF parser for researchers and developers who need scientific papers as usable text, including their equations and tables. It converts academic PDFs into Markdown-style documents with LaTeX notation, so the output retains structure that plain text extraction can lose.

The Python tool runs on your own machine and supports CPU or GPU processing, including GPU use on Windows. It provides a command-line interface for individual PDFs or batches and a local API for applications that need document conversion. You can process selected pages instead of a whole paper. Model checkpoints are available to download, with small and base variants.

Its output uses .mmd files, a markup format mostly compatible with Mathpix Markdown, with LaTeX tables. Markdown compatibility processing is enabled by default. This makes Nougat relevant for document-processing projects where mathematical content needs to survive conversion alongside the surrounding prose.

Document type matters. Nougat was trained on scientific papers from arXiv and PMC and works best with English papers. Other languages using Latin scripts may work, but Chinese, Russian and Japanese aren't supported. Its failure detection can also mark pages as missing on some CPUs or older GPUs.

The project builds on Donut and includes tools for preparing training datasets, fine-tuning models and evaluating results. The code is open source under MIT; the model weights carry a separate CC-BY-NC license that restricts commercial use.