Compare
pdfboss vs PyMuPDF.
Both read PDFs from Python. PyMuPDF wraps MuPDF, a C engine, and ships under the AGPL-3.0 or a commercial license from Artifex. pdfboss is written from scratch in Rust, dual-licensed MIT OR Apache-2.0, and faster on the operations measured below. Here is where each one stands, gaps included.
| pdfboss | PyMuPDF | |
|---|---|---|
| License | MIT OR Apache-2.0 | AGPL-3.0, or a commercial license from Artifex |
| Engine | Written from scratch in safe Rust; no C dependencies | Python bindings to MuPDF, written in C |
| Text extraction | 10,312 pages/s | 435 pages/s (40 real-world PDFs, same machine) |
| Rendering to PNG, PPM, BMP or JPEG | 161.9 pages/s | 89.4 pages/s (38 certified files, 888 pages) |
| Scanned JBIG2 book | 65.2 pages/s | 51.7 pages/s (544 pages at 1:1) |
| PDF to Markdown | pdfboss md: NID 0.883, 0.15 s for 200 PDFs | pymupdf4llm: NID 0.886, 17.12 s for 200 PDFs |
| Async and remote | AsyncDocument over files and HTTP range requests | Synchronous API |
| Editing existing PDFs | Not supported; creation only | Annotations, forms, redaction, page manipulation |
| OCR | No | Yes, through Tesseract |
| Other formats | PDF only | XPS, EPUB, MOBI, FB2, CBZ, SVG and images |
| Reporting | Every dropped or approximated item is reported per page | Warnings through MuPDF's error stack |
Throughput rows come from the pdfboss benchmark harness; the Markdown row from opendataloader-bench, where the pymupdf4llm timing is the one published with the corpus. License statements are PyMuPDF’s own.
Migration
The reading path, side by side.
Documents index like lists, pages are 0-based in both, and render returns PNG, PPM, BMP or JPEG bytes directly instead of a pixmap you serialize. extract_text() on the document covers all pages and spreads them across cores.
before.py · PyMuPDF
import pymupdf
doc = pymupdf.open("report.pdf")
text = "".join(page.get_text() for page in doc)
pix = doc[0].get_pixmap(matrix=pymupdf.Matrix(2, 2))
png = pix.tobytes("png")
for info in doc[0].get_images(full=True):
image = doc.extract_image(info[0])after.py · pdfboss
import pdfboss
doc = pdfboss.Document("report.pdf")
text = doc.extract_text()
png = doc[0].render(scale=2.0)
for image in doc[0].extract_images():
image.width, image.height, image.dataWhere PyMuPDF is ahead
What PyMuPDF does that pdfboss does not.
- Edits existing PDFs: annotations, form fields, redaction, page insertion and deletion, incremental saves.
- OCR through Tesseract, and rendering of many non-PDF formats.
- Two decades of MuPDF rendering coverage on unusual files. pdfboss is younger; it reports what it cannot paint rather than guessing, and its rendering benchmark is restricted to files it provably rasterizes completely.
Where pdfboss is ahead
What pdfboss does that PyMuPDF does not.
- Permissive licensing: MIT OR Apache-2.0, so commercial and SaaS use carries no copyleft obligation and no commercial license to buy.
- Speed on the read path: 24 times faster text extraction and 1.8 times faster rendering on the measured corpus, with no C in the wheel.
- Async reading over HTTP range requests, so a page from a remote file arrives without downloading the file.
- A jq-style query language and a terminal explorer over the document's element tree (pdfboss q, hex, json, tui).
- Deterministic PDF creation from a composable Python API and a Markdown-to-PDF path with CSS themes.
Questions
- Is pdfboss a drop-in replacement for PyMuPDF?
- No. The APIs differ and pdfboss covers a smaller surface: reading PDFs (text, Markdown, styled spans, rendering, embedded images) and creating new ones, but not editing existing files. Migrating the reading path is usually a few lines per call site.
- Can I use pdfboss in commercial or closed-source software?
- Yes. pdfboss is dual-licensed MIT OR Apache-2.0, permissive licenses with attribution requirements and no copyleft. PyMuPDF is licensed under the AGPL-3.0 unless you buy a commercial license from Artifex.
- Is pdfboss faster than PyMuPDF?
- On the operations measured, yes: text extraction runs about 24 times faster over 40 real-world PDFs (10,312 against 435 pages per second), rendering about 1.8 times faster over 888 certified pages (161.9 against 89.4), and a 544-page JBIG2 scan about 26% faster (65.2 against 51.7). Converting the 200-document opendataloader-bench corpus to Markdown takes 0.15 seconds against pymupdf4llm's 17.12, with a slightly lower reading-order score (0.883 against 0.886).
- Does pdfboss need MuPDF or any other C library installed?
- No. pdfboss is written from scratch in Rust, including its JPEG 2000, JBIG2, CCITT and ICC decoders, and the Python wheel has no native dependency beyond CPython 3.12 or later.