B.L.A.S.T. OCR vs PyMuPDF4LLM: Scanned-PDF OCR vs Text-Layer Extraction
PyMuPDF4LLM and B.L.A.S.T. OCR solve different halves of the same problem. PyMuPDF4LLM extracts Markdown directly from a PDF's existing text layer with no ML models or GPU, and only runs OCR as an opt-in plugin on pages that lack one. B.L.A.S.T. is built OCR-first for scanned and image-only documents, running ONNX inference by default with table structure, formula parsing, and MCP integration.
Executive summary
They have not been run head-to-head on the same corpus in this project's eval harness, so no throughput or CER comparison is made here. The table below compares documented architecture and features only.
| Dimension | PyMuPDF4LLM (Artifex) | B.L.A.S.T. OCR Engine |
|---|---|---|
| Primary design target | Born-digital PDFs with an existing text layer | Scanned/image PDFs with no text layer |
| OCR behavior | Opt-in plugin, skipped by default when text exists | Default engine (RapidOCR/ONNX) on every page |
| GPU/ML for base extraction | None (default text-layer path) | ONNX Runtime required (CPU/CUDA/DirectML) |
| Table structure extraction | Layout-based Markdown tables | Markdown/HTML, scored with a built-in TEDS evaluator |
| Formula/LaTeX parsing | Unicode math symbols only, relies on downstream LLM to interpret as LaTeX | Built-in LaTeX ($...$, $$...$$) to KaTeX conversion |
| Native MCP server | Yes — pymupdf4llm-mcp, a companion package from the PyMuPDF/Artifex team | Yes — built into the core library (blast_ocr/mcp_server.py) |
| License of own source | AGPL-3.0 (commercial license via Artifex) | MIT |
License nuance (read before you pick a side)
B.L.A.S.T.'s own source is 100% MIT. But B.L.A.S.T. ships an optional pdf extra (pyproject.toml: pdf = ["pymupdf>=1.23.0"]) used for dual-layer sandwich PDF generation and PDF page streaming. That extra depends on the base pymupdf (fitz) library — the same AGPL-3.0/commercial-dual-licensed engine underneath PyMuPDF4LLM.
This is not a case where B.L.A.S.T. avoids that dependency entirely. B.L.A.S.T.'s core OCR pipeline does not require pymupdf at all — it's imported lazily, only inside the sandwich-PDF and streaming code paths, and only when that optional extra is installed — whereas PyMuPDF4LLM's core text extraction is PyMuPDF, so the AGPL obligation is unavoidable there regardless of which features you use. If your deployment needs to stay clear of AGPL entirely, check whether you're installing B.L.A.S.T.'s pdf extra — the base install without it does not pull in pymupdf.
Bottom line: who should choose what?
- Choose PyMuPDF4LLM if: your PDFs are mostly born-digital (exported from Word, LaTeX, or a web page) and already have a real text layer. You get Markdown extraction with zero ML overhead and no GPU, with OCR only as a fallback for the occasional scanned page.
- Choose B.L.A.S.T. if: your source documents are scanned paper, faxes, or flat images with no text layer at all, and you also want table structure extraction, LaTeX formula parsing, a native MCP server, and a permissively-licensed (MIT) core engine.
Full writeup, including the JSON-LD source and author box, is at docs/comparisons/blast-vs-pymupdf4llm.md.