Best Open-Source PDF-to-Markdown Tools in 2026: Marker vs Docling vs MinerU vs pdf-craft vs PyMuPDF
Youβre building a RAG pipeline, processing research papers, or digitizing a bookshelf β and you need PDFs turned into clean Markdown. The good news: open-source tooling for this has gotten very good. The bad news: there are now enough options that picking the right one takes real research.
I did that research. Here are the five tools worth knowing about in 2026, what each does best, and which one you should actually use.
Updated: November 2026. In October 2026, Artifex (the company behind PyMuPDF) shipped PyMuPDF 2.0, folding PyMuPDF4LLM and PyMuPDF Layout into the core package and switching on GNN-based layout analysis by default. Thatβs enough of a change to move PyMuPDF from βfast but basicβ to a genuine Marker/Docling competitor for CPU-only pipelines β see the updated entry and comparison table below.
The Contenders
1. Marker (datalab-to/marker) β β 19K+
The all-rounder. Marker converts PDFs, images, DOCX, PPTX, XLSX, HTML, and EPUB to Markdown, JSON, or chunks. It uses Surya OCR for text recognition and supports GPU, CPU, and Apple MPS.
What sets Marker apart is the optional --use_llm flag that layers an LLM on top for accuracy-critical documents. Without it, you get fast local conversion. With it, you get near-perfect output on messy layouts.
Best for: General-purpose conversion, multi-format pipelines, teams that want one tool for everything.
Strengths:
- Widest format support (not just PDFs)
- Optional LLM boost for higher accuracy
- Built-in benchmarking suite
- Structured JSON extraction with schema support
- ~25 pages/sec throughput on H100 in batch mode
Trade-offs:
- LLM mode adds latency and API cost
- Surya OCR is good but not best-in-class for every language
2. Docling (docling-project/docling) β β 20K+
The enterprise pick. Built by IBM Research, Docling is designed for production RAG pipelines. It outputs a structured DoclingDocument that preserves semantic hierarchy β not just text, but the meaning of the document structure.
Docling handles PDF, DOCX, PPTX, XLSX, HTML, images, audio (WAV/MP3), LaTeX, and plain text. It has first-class integrations with LlamaIndex, LangChain, and other gen AI frameworks.
Best for: Enterprise document processing, structured extraction, teams already in the LlamaIndex/LangChain ecosystem.
Strengths:
- Rich structured output (DoclingDocument format)
- Advanced layout analysis and reading order detection
- Native integrations with RAG frameworks
- Handles audio and image formats too
- Linux Foundation AI project β strong governance
Trade-offs:
- Heavier setup than simpler tools
- CPU mode is notably slower
- Formula support less mature than Marker
3. MinerU (opendatalab/MinerU) β β 30K+
The heavy hitter. MinerU has the most GitHub stars for a reason β itβs extremely capable at complex document layouts, especially CJK content. Built by OpenDataLab (Shanghai AI Lab), it uses PaddleOCR and custom layout detection models.
MinerU outputs both Markdown and JSON, with explicit support for agentic workflows. It also has the broadest hardware support of any tool here, running on NVIDIA, AMD, and a dozen Chinese accelerator platforms (Ascend, Kunlunxin, Cambricon, etc.).
Best for: CJK documents, complex academic papers, teams with diverse GPU hardware.
Strengths:
- Best layout detection for complex documents
- Excellent CJK language support
- Runs on practically any accelerator
- Active development (v2.7.6 as of Feb 2026)
- LLM-ready JSON output
Trade-offs:
- Heavier dependencies (PaddlePaddle ecosystem)
- Setup can be complex on non-standard hardware
- Primarily focused on PDFs (not multi-format)
4. pdf-craft (oomol-lab/pdf-craft) β β 4.6K
The book specialist. pdf-craft is purpose-built for converting scanned books to Markdown or EPUB. It uses DeepSeek OCR and runs entirely locally β no network requests, no API keys, no cloud dependency.
Starting with v1.0, pdf-craft dropped LLM-based text correction entirely in favor of pure OCR. The result is faster, more predictable conversion. It automatically generates EPUB table of contents and handles footnotes, formulas, and tables.
Best for: Digitizing physical books, library/archive projects, anyone who wants fully offline conversion.
Strengths:
- Purpose-built for scanned books (headers/footers auto-filtered)
- EPUB output with auto-generated TOC
- Fully local β zero network calls
- DeepSeek OCR handles tables and formulas well
- Clean, focused codebase
Trade-offs:
- PDF-only (no DOCX, PPTX, etc.)
- Requires GPU for practical use
- Smaller community than Marker/Docling/MinerU
- No LLM correction option (removed in v1.0)
5. PyMuPDF (pymupdf/PyMuPDF) β β 10K+ β Updated for 2.0 (October 2026)
The consolidated option. Through 2025 and early 2026, getting LLM-ready Markdown out of PyMuPDF meant installing three separately versioned packages β PyMuPDF for the engine, PyMuPDF4LLM for structured extraction, and PyMuPDF Layout for reading order and table/header detection β and remembering to import Layout before anything else or results would silently degrade.
PyMuPDF 2.0 (released October 19, 2026) ends that. PyMuPDF4LLM and PyMuPDF Layout are now part of the core package. One pip install pymupdf, one import, and layout-aware Markdown/JSON/plain-text output is on by default.
The bigger change is under the hood: PyMuPDFβs layout engine works directly on a PDFβs internal document structure using graph neural networks, rather than rendering each page to an image and running a vision model over it like most layout tools do. Thatβs how it gets reading order, multi-column handling, and header/footer detection right while staying CPU-only. It also does selective OCR now β pages are inspected first, and only the regions that actually need OCR get it, with the recognized text merged back into the existing digital content.
If you already have import pymupdf4llm in your codebase, it keeps working as a compatibility shim β nothing breaks on upgrade.
Best for: Native and lightly-scanned PDFs, CPU-only environments, teams who want Marker/Docling-grade layout understanding without a GPU or three packages to version-pin.
Strengths:
- GNN-based layout analysis on by default β reading order, headers/footers, tables, multi-column pages
- Selective/hybrid OCR β only runs where itβs needed, merges with native text
- Markdown, JSON (with bounding boxes + element types), and plain-text output built in
- LlamaIndex and LangChain integrations included
- Still CPU-only, still fast, no API keys, no cloud calls β works air-gapped
- One package instead of three; old
pymupdf4llmimports still work
Trade-offs:
- Formula recognition still isnβt a strength relative to Marker or pdf-craft
- CJK layout handling is improved but not at MinerUβs level
- Newer consolidated release β less time in production than Marker/Docling/MinerUβs current versions
Head-to-Head Comparison
| Feature | Marker | Docling | MinerU | pdf-craft | PyMuPDF 2.0 |
|---|---|---|---|---|---|
| GitHub Stars | 19K+ | 20K+ | 30K+ | 4.6K | 10K+ |
| OCR Engine | Surya | Custom | PaddleOCR | DeepSeek | Selective hybrid OCR |
| GPU Required | No (optional) | No (optional) | Recommended | Yes | No |
| Scanned PDFs | β | β | β | β (specialized) | β (selective) |
| Multi-format | PDF, DOCX, PPTX, EPUB, images | PDF, DOCX, PPTX, HTML, audio | PDF, Office files | ||
| Tables | β β | β β | β β β | β | β β |
| Formulas | β β | β | β β | β β | β |
| EPUB Output | β | β | β | β | β |
| LLM Boost | Optional | β | β | β (removed) | β |
| Fully Local | β | β | β | β | β |
| CJK Support | Good | Good | Excellent | Good | Improving |
| RAG Integration | JSON/chunks | LlamaIndex, LangChain | JSON | β | LlamaIndex, LangChain, JSON w/ bounding boxes |
| Layout method | Vision model (Surya) | Custom ML | Vision model (PaddleOCR) | Vision model (DeepSeek OCR) | Graph neural network on doc structure |
| Speed (native PDF) | Fast | Moderate | Moderate | N/A | Fastest |
| Speed (scanned) | Fast (GPU) | Moderate | Fast (GPU) | Fast (GPU) | Fast (CPU, selective OCR) |
Decision Tree
Start here:
-
Is your PDF native (has selectable text), or only lightly scanned? β Try PyMuPDF 2.0 first. Itβs fast, CPU-only, and its GNN layout engine now genuinely competes with Marker/Docling on reading order and structure β not just a fallback anymore. If the output quality isnβt there, move to Marker.
-
Are you processing scanned books? β pdf-craft is built for exactly this. If you also need multi-format support, use Marker instead.
-
Do you need structured output for a RAG pipeline? β Docling if youβre in the LlamaIndex/LangChain ecosystem. Marker if you want JSON schema extraction. PyMuPDF 2.0 is now a legitimate third option here too, with the same framework integrations and lower resource requirements.
-
Complex academic papers with CJK text? β MinerU. Nothing else comes close for Chinese/Japanese/Korean layout detection.
-
Want one tool that does everything reasonably well? β Marker remains the Swiss Army knife for multi-format pipelines. PyMuPDF 2.0 is the pick if you want that breadth without a GPU.
The Bottom Line
The PDF-to-Markdown space has matured fast. Two years ago you were choosing between βbadβ and βless bad.β Now youβre choosing between genuinely good tools with different specializations β and the consolidation of PyMuPDF 2.0 closes what used to be the gap between βquick and dirtyβ and βreal layout understanding.β
If youβre only going to install one tool: Marker is still the safest default for multi-format, GPU-friendly pipelines. If you want the same caliber of layout analysis without a GPU or extra packages, PyMuPDF 2.0 is now the better starting point than the old PyMuPDF4LLM was. For specific use cases β books (pdf-craft), enterprise RAG (Docling), CJK (MinerU) β pick the specialist.
All five are open source, all run locally, and all are actively maintained. The real winner is the ecosystem.
Have a favorite PDF-to-Markdown tool I missed? Let me know.