mineru vs opendataloader-pdf
Two of the top projects in Data and scraping for AI, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
mineru
Document parser that converts PDFs, images and Office files to Markdown and JSON
opendataloader-pdf
PDF parser that outputs Markdown, JSON with bounding boxes, and Tagged PDF
Verdict
mineru ranks higher, 94 to 87. Of the 4 score parts we measure for both, mineru leads on 2, opendataloader-pdf on 1 and 1 is even.
- Adoptionmineru leads by 22 points, 90 to 68: a wide gap.
- FreshnessEven at 100.
- Maintenancemineru leads by 5 points, 99 to 94.
- Agent-readyopendataloader-pdf leads by 15 points, 85 to 70: a wide gap.
Pick mineru if you are
Engineers feeding PDFs and Office documents to LLM pipelines or agents
- Handles scanned PDFs, tables and formulas, with four quality tiers from fast preview to advanced
- Stable doc/tier/page/block locators allow paged reading and citation of long documents
Pick opendataloader-pdf if you are
Engineers feeding PDFs into RAG pipelines or automating PDF tagging
- Bounding boxes and element types for every extracted item, usable for source citations
- Local mode needs no GPU; README reports 0.015 s/page
| What we compare | mineru | opendataloader-pdf |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 90, widely used | 68, popular |
| Freshness | 100, active | 100, active |
| Maintenance | 99, healthy | 94, healthy |
| Agent-ready | 70, partly | 85, ready |
| Facts from GitHub and the README | ||
| Stars | 82.9k | 30.1k |
| License | custom license (read the license) | Apache-2.0 (permissive) |
| Last commit | Oct 2026 | Oct 2026 |
| Last release | Oct 2026 | Oct 2026 |
| Language | Python | Java |
| Docker | Yes | No |
| GPU | Optional | Not needed |
| arm64 or Apple Silicon | Mentioned | Not stated |
mineru
MinerU parses PDFs, scanned images, Office, OpenDocument, EPUB, OFD, HTML and CSV files into Markdown, HTML, LaTeX, DOCX and structured content lists. It offers four parsing tiers (flash, basic, standard, advanced), a local document library with search, and page/block locators for citations. It ships a CLI, Python SDK, API, router and Gradio WebUI, plus an agent skill.
Who it is for: Engineers feeding PDFs and Office documents to LLM pipelines or agents
Strengths
- Handles scanned PDFs, tables and formulas, with four quality tiers from fast preview to advanced
- Stable doc/tier/page/block locators allow paged reading and citation of long documents
- Small models run on ONNX CPU; the VLM can run on llama.cpp, vLLM or LMDeploy
- Documents are not uploaded to the official service unless remote parsing is configured
Weaknesses
- Office, HTML, EPUB, OFD and CSV files only get flash-tier parsing; higher tiers are PDF and image only
- Docker deployment for non-NVIDIA devices is pending an update
- Highest throughput needs an NVIDIA GPU and mineru[full]; Windows needs torch installed separately
- License is reported as NOASSERTION, so terms need checking before use
- GPU optional
- Docker + Compose
- Needs Python >=3.10,<3.15, llama.cpp, vLLM or LMDeploy (VLM runtime), ONNX or Torch
- Models: MinerU VLM, ONNX small models, Torch small models, llama.cpp, vLLM
opendataloader-pdf
Java-based PDF parser with Python, Node.js and Java SDKs. It extracts Markdown, JSON with per-element bounding boxes, and HTML, using XY-Cut++ reading order in a deterministic local mode. An optional hybrid mode sends complex pages to a Docling-based backend for OCR, borderless tables, LaTeX formulas and chart descriptions. It also auto-tags untagged PDFs into Tagged PDF.
Who it is for: Engineers feeding PDFs into RAG pipelines or automating PDF tagging
Strengths
- Bounding boxes and element types for every extracted item, usable for source citations
- Local mode needs no GPU; README reports 0.015 s/page
- Hybrid mode adds OCR, LaTeX formulas and chart descriptions; OCR engine is selectable
- Reports 0.907 overall on its own 200-PDF benchmark, ahead of docling and marker
Weaknesses
- Requires Java 11+ on the host, even when using the Python package
- Each convert() call spawns a JVM, so calls must be batched
- PDF/UA export and accessibility studio are enterprise-only
- Does not process Word, Excel or PowerPoint; benchmark is run by the project itself
- no GPU
- Needs Java 11+, Python 3.10+ (Python SDK), Docling backend (hybrid mode)
- Models: Docling, EasyOCR, RapidOCR, Tesseract, SmolVLM (256M)
- port 5002
More in Data and scraping for AI
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.