MinerU is hard to avoid in document parsing these days: it converts layout-heavy PDFs into clean Markdown or JSON, turns formulas into LaTeX and tables into HTML, and routes scans through OCR automatically — exactly what LLM training and RAG knowledge bases need in high-quality corpora. It has passed 50,000 stars on GitHub, but stars only measure attention. Whether it deserves a place in your stack depends on the accounts below.
Official repository
- Platform: GitHub
- Organization: opendatalab (the open-source community under Shanghai AI Laboratory)
- Project name: MinerU
It was born inside InternLM's pre-training data pipeline, originally to fix symbol and formula conversion in scientific literature, and later became a general-purpose document parser.
Where it is strong
Ordinary PDF extractors chop text block by block, so two-column papers come out scrambled and formulas turn to gibberish. MinerU first recognizes structure with layout analysis models: headers, footers, and page numbers are removed, content is reordered to human reading order, and heading levels, paragraphs, and lists are preserved. The project offers two routes: the pipeline route chains small specialist models — fast and controllable; the VLM route uses a dedicated document-understanding vision model to read whole pages, with better accuracy on complex layouts and cross-page tables. Output includes coordinate-bearing intermediate JSON besides Markdown, which helps secondary development.
Deployment cost: not light
Installation is one pip command, but a working setup downloads a full set of model weights, and first-time configuration is not beginner-friendly. The pipeline route runs on CPU alone, just slowly — batch-processing hundreds of pages takes patience. For usable speed, a GPU with 8GB or more VRAM is the practical bar, and the VLM route asks for more. Windows brings more dependency trouble than Linux; teams usually deploy with Docker or on a Linux server. Docling deploys more lightly by comparison, but for complex Chinese layouts and formulas, MinerU is currently the better fit.
The real caveats
First, the license changed. The project started under AGPLv3 and, from the 3.1.0 release in 2026, moved to the MinerU Open Source License based on Apache 2.0, with attribution requirements and commercial-use thresholds. Teams embedding it into commercial products or public services should have legal review the current license text rather than rely on a vague "old open source, use freely" impression — and model weights from different versions may carry different terms, so check each download.
Second, accuracy has limits. Poor scans, oddly laid-out tables, and handwriting can trip up even the VLM route, and its mistakes look convincing. Always sample-check output before it enters a RAG index; never trust the pipeline blindly.
Third, it iterates fast. Releases are frequent and parameters and output structures have changed repeatedly. Pinning versions and keeping regression samples before upgrading is a required discipline for production pipelines.
Who should adopt it, who should pass
If you batch-process papers, financial reports, or manuals to feed large models, and your team can maintain a Python environment, MinerU is a first-tier open-source choice today. If you only convert a PDF once in a while, the official online version or a ready-made converter is the better deal — building a deployment for a single conversion never pays back the time.
