MinerU: 70.6K Stars — Convert Any Document to LLM-Ready Markdown

Alpaca Trading API 2026: The Commission-Free Stock Brokerage API for Algorithmic Trading — Setup Guide • GraphRAG: Microsoft’’s Graph-Based RAG for Better LLM Answers (33K Stars) — Practical 2026 Guide MinerU — the open-source document parsing engine that turned 70,600 GitHub stars in just over a year.

In the age of AI coding agents and RAG pipelines, the biggest bottleneck isn’t model capability — it’s data quality. You can have the most powerful LLM in the world, but if your input documents are messy PDFs with broken tables, unreadable formulas, and garbled headers, the output will be garbage.

Enter MinerU, a document parsing tool from OpenDataLab (the team behind InternLM) that converts PDF, DOCX, PPTX, XLSX, images, and web pages into clean, structured Markdown and JSON — specifically designed for downstream LLM, RAG, and Agent workflows.

With 70,600+ GitHub stars, 5,900 forks, and 960 stars gained today alone, MinerU has quickly become the go-to open-source solution for document-to-LLM-pipeline conversion.

What Is MinerU? #

MinerU was born during the pre-training process of InternLM, a Chinese large language model. The team noticed that converting scientific papers and technical documents into machine-readable format was a massive pain point — especially for formulas, tables, and complex layouts.

So they built MinerU.

Unlike traditional PDF parsers that just extract raw text, MinerU understands document structure:

• Removes headers, footers, footnotes, and page numbers that break semantic coherence
• Preserves reading order for single-column, multi-column, and complex layouts
• Converts formulas to LaTeX, tables to HTML
• Detects scanned and garbled PDFs and automatically enables OCR
• Supports 109 languages for OCR recognition
• Outputs multimodal Markdown, NLP Markdown, and JSON sorted by reading order

The result is document content that AI agents and LLMs can actually understand.

Installation and Setup #

MinerU offers multiple installation paths depending on your needs:

pip (Recommended for most users) #

pip install mineru

Docker #

docker pull mineru/mineru:latest
docker run --gpus all -v $(pwd):/data mineru/mineru:latest

Local Development #

git clone https://github.com/opendatalab/MinerU.git
cd MinerU
pip install -e .

MinerU supports both CPU-only and GPU-accelerated inference. For GPU acceleration, install with CUDA support:

pip install mineru[cuda]

On macOS with Apple Silicon, MinerU leverages MPS (Metal Performance Shaders) for acceleration.

Three Parsing Engines #

MinerU provides three different parsing backends, each optimized for different scenarios:

1. Pipeline Backend (Fast & Stable) #

The pipeline backend is the default choice for most users. It’s fast, stable, and produces no hallucinations. It runs efficiently on CPU and is ideal for batch processing.

mineru ./input.pdf -o ./output/

Best for: High-volume document processing, CI/CD pipelines, CPU-only environments.

2. VLM Engine (High Accuracy) #

The VLM (Vision Language Model) engine uses MinerU’s proprietary MinerU2.5-Pro-2604-1.2B model for state-of-the-art parsing accuracy. It excels on complex documents with mixed layouts, handwritten text, and dense formulas.

mineru ./complex.pdf -o ./output/ --engine vlm-engine

Best for: Complex scientific papers, scanned documents, handwritten content, multi-language OCR.

3. Hybrid Engine (Balanced) #

The hybrid engine combines native text extraction with VLM-based analysis. Starting from version 3.3, it includes an effort parameter with medium and high levels:

• Medium effort: 35-220% faster than high, with only 0.13-point accuracy drop on OmniDocBench
• High effort: Maximum accuracy with image analysis support

mineru ./document.pdf -o ./output/ --engine hybrid-engine --effort medium

Best for: Production workloads where you need to balance speed and accuracy.

Supported Formats #

MinerU supports a comprehensive range of input formats:

| Format | Support Level | Notes | |