77,975 stars · MinerU Open Source License — Apache 2.0 plus three added conditions, read from the LICENSE file: a separate commercial licence above 100 million monthly active users or 20 million USD monthly revenue, prominent attribution for any online service built on it, and automatic termination if either is breached. NOT plain Apache-2.0. · mineru-4.0.7-released (2026-09-23), read from /releases/latest and confirmed by ungh · Track this in Scout
It converts PDFs into clean Markdown or JSON, keeping headings, tables, equations and reading order.
▶Repo detailsthe review · specs · pros & cons · install
What it is
MinerU is a document extraction pipeline. It works out the layout of each page, reads text that exists only as an image, and rebuilds the document with its headings, tables, formulas and reading order intact. It ships a command, a local web page and a server mode.
What it is good for. Any pile of PDFs somebody needs to search, quote or feed into another program: scanned contracts, research papers, manuals, statements, old invoices. It is unusually strong on documents full of equations and on Chinese-language text, where simpler tools scramble the order.
- It handles the hard cases: multi-column pages, formulas, tables that run across pages, and pages that are nothing but an image.
- Very actively developed. The release tagged
mineru-4.0.7-releasedwas published on 23 September 2026, two days before this edition. - There is more than one way in. A command, a local web page and a server mode mean it fits both a one-off job and a pipeline.
- The licence is not plain Apache-2.0, whatever a summary says. The LICENSE file is the MinerU Open Source License: Apache 2.0 with three additions. A separate commercial licence is required above 100 million monthly active users or 20 million US dollars of monthly revenue. Any online service built on it must state prominently that MinerU is used. Breaking either condition ends the licence automatically. For almost everybody the money thresholds are irrelevant. The attribution requirement is not.
- It is heavy. The project's own requirements ask for at least 16 GB of memory, 32 GB recommended, and about 20 GB of disk for the model files. Check the machine has room before starting.
- It converts and stops. It does not store, index or search the results, so something else has to hold them afterwards.
docling-project/docling67,760 stars, covered in Edition 3; it reads Word and PowerPoint files as well and is lighter on memory, at some cost in accuracy on equations.
Track this in Scout
datalab-to/marker39,918 stars, covered in Edition 9; plain Apache-2.0 with no thresholds and no attribution clause, and it is more comfortable on a machine without a graphics card.
Track this in Scout- getomni-ai/zerox
12,259 stars, converts by sending each page to a vision model instead of running local layout models, so it costs money per page; its last code landed on 20 May 2025.
Track this in Scout
python3 -m venv venv source venv/bin/activate pip install -U "mineru[core]" # one file mineru -p report.pdf -o ./out # a whole folder mineru -p ./pdfs -o ./out
