Markdown to PDF Comparison 2026 10 13

Markdown-to-PDF Comparison 2026-10-13

RayoMD 4.3.1 against glu, Qt 6's rich-text printing, Pandoc with Typst, WeasyPrint or XeLaTeX, and md-to-pdf, converting the repository's tester.md, each tool as a fresh process, start to exit. The tools render different Markdown subsets and look different, so this compares time to a finished PDF, memory and install size, not identical output. The page renders below show what each one makes of the same document.

Setup

• AMD Ryzen 5 PRO 4650GE (6 cores, 12 threads), Linux 7.0, an Arch Linux Docker container (archlinux:latest of 2026-10-12) with --network none, files on tmpfs.
• Four pinned cores (--cpuset-cpus=3,5,9,11) for the headline, one (--cpuset-cpus=3) beside it; nothing else heavy running.
• Median of 20 runs after one uncounted warm-up, every output checked to be a complete PDF (qpdf --check, page count, a phrase found by pdftotext).
• tools/benchmark.py competitors (scripts/compare_markdown_pdf_tools.py), each command started in the document's folder. Its timing no longer waits with a timeout, which polled and added about 1.1 ms to every run on one core (commit f11b5c4); earlier reports were not affected on more than one core.

Tool	Version	Command
RayoMD	4.3.1 (release rayomd-4.3.1-linux-x64)	rayomd --export IN OUT native modern normal --no-image-cache
glu	0.0.46 (release glu-linux-amd64)	glu -q -o OUT IN
Qt 6	qt6-base 6.12.0	a 110-line program: QTextDocument::setMarkdown, then print() to a QPrinter(HighResolution) with PDF output; IBM Plex Mono at 11 pt, QT_QPA_PLATFORM=offscreen
Pandoc + Typst	Pandoc 3.11 (pandoc-cli), Typst 0.15.1	pandoc IN -o OUT --pdf-engine=typst …
Pandoc + WeasyPrint	Pandoc 3.11, WeasyPrint 70.0	pandoc IN -o OUT -t html --pdf-engine=weasyprint …
md-to-pdf	5.2.5, Node 26.11.1, Chromium 154	scripts/md_to_pdf_runner.cjs (A4, 20 mm margins)
Pandoc + XeLaTeX	Pandoc 3.11, TeX Live 2026	pandoc IN -o OUT --pdf-engine=xelatex …

The Pandoc routes read -f markdown+tex_math_single_backslash-raw_tex (so \[ … \] is display math, as in RayoMD) with DejaVu Sans and DejaVu Sans Mono; md-to-pdf ran Chromium with --no-sandbox (inside the container).

Documents

The same two documents as in Markdown-to-PDF Speed Comparison 2026-10-12: tester.md without the lines that test fallbacks (two remote images, a missing local image and a formula with an unknown TeX command, on which glu and XeLaTeX stop with an error and Pandoc waits on the network).

Document	Bytes	SHA-256 (first 16)	Contents
tester-local.md	4,880	4833c4ff1c15fe54	keeps the 1024×1024 RGBA mascot PNG
tester-noimg.md	4,750	0d35d7a801d21274	without images

Four cores

Tool	Without images	× RayoMD	With the PNG	× RayoMD	Pages	Page size	Formulas
RayoMD	1.48 ms	1×	7.75 ms	1×	5 / 6	A4	typeset
glu	54.49 ms	37×	84.82 ms	11×	3 / 9	A4	TeX source
Qt 6	69.21 ms	47×	111.79 ms	14×	3 / 6	A4	TeX source
Pandoc + Typst	345.62 ms	234×	375.99 ms	49×	4 / 6	Letter	typeset
Pandoc + WeasyPrint	967.80 ms	654×	1,238.19 ms	160×	4 / 6	A4	MathML text and TeX source
md-to-pdf	1,804.97 ms	1,220×	1,902.96 ms	246×	3 / 6	A4	TeX source
Pandoc + XeLaTeX	2,096.73 ms	1,417×	3,351.33 ms	432×	4 / 6	Letter	typeset

One core

Tool	Without images	× RayoMD	With the PNG	× RayoMD
RayoMD	1.52 ms	1×	11.29 ms	1×
glu	63.96 ms	42×	98.00 ms	8.7×
Qt 6	69.03 ms	45×	112.42 ms	10×
Pandoc + Typst	336.11 ms	221×	375.89 ms	33×
Pandoc + WeasyPrint	975.93 ms	642×	1,246.85 ms	110×
Pandoc + XeLaTeX	2,162.78 ms	1,423×	3,458.34 ms	306×
md-to-pdf	2,254.66 ms	1,483×	2,352.88 ms	208×

A second core helps RayoMD only with the PNG, which it re-encodes on up to four cores; it helps md-to-pdf (about 450 ms), XeLaTeX (about 100 ms) and glu (10 to 13 ms), and not Pandoc + Typst, Qt or WeasyPrint.

Memory and install size

Peak RSS of tester-noimg.md (/usr/bin/time %M, median of 5). For the tools that run several processes this is the largest one, not the sum.

Tool	Peak RSS	Installed size beyond the Arch base image (approximate)
RayoMD (static edition in brackets)	3.9 MiB (2.5 MiB)	1.2 MiB (2.1 MiB)
glu	33.2 MiB	25 MiB
Qt 6	38.5 MiB	373 MiB (comes with any Qt desktop)
md-to-pdf	205.2 MiB	about 1.0 GiB (Node.js, npm, 39 MiB of modules, Chromium)
Pandoc + Typst	239.3 MiB	about 0.65 GiB
Pandoc + WeasyPrint	240.5 MiB	about 0.9 GiB
Pandoc + XeLaTeX	339.9 MiB	about 1.4 GiB

Install sizes are pacman -Qi "Installed Size" over each package's dependency closure, without the 137 packages of the base image; the closures overlap.

Output

• RayoMD, Typst and XeLaTeX typeset the formulas; glu, Qt and md-to-pdf print their TeX source; WeasyPrint prints MathML text followed by the TeX annotation.
• Qt prints \pagebreak literally and does not scale the 1024-pixel PNG to the page, so it is clipped; glu lays the PNG document out on 9 pages.
• Every font is embedded except RayoMD's, which uses the PDF standard fonts where the text allows it, and for formulas Times and Symbol. In this container, without gsfonts, Poppler has no Symbol substitute and draws ∫ ∑ ≤ as nothing (visible in RayoMD's first page above); with gsfonts, as most desktops have, they render. The PDFs are valid either way.
• Typography: Typst and XeLaTeX set justified, hyphenated text with fine spacing, which none of the fast tools do.

Against 2026-10-12

That comparison ran on Debian 12 without pinning, with Pandoc 3.12.1's static build and TeX Live 2022. Pandoc + Typst took 112 ms there and 346 ms here: Arch's pandoc-cli is linked against Arch's Haskell library packages and takes 113 ms for pandoc --version alone, and typst compile took 203 ms on its own (median of 10). md-to-pdf went from 1,670 to 1,805 ms. RayoMD 4.0.0 took 1.6 ms there; 4.3.1 takes 1.5 ms.

RayoMD 4.4.0

The same day, the same container and harness, only RayoMD measured again with the 4.4.0 release binary (median of 20): without images 1.21 ms on four cores and 1.25 ms on one (4.3.1 measured 1.39 and 1.45 ms in that run), with the PNG 7.47 and 10.99 ms; peak memory 3.4 MiB (4.3.1: 3.9 MiB). Against the other tools' figures above, that is 45× faster than glu, 57× than Qt 6, 286× than Pandoc + Typst and 1,733× than Pandoc + XeLaTeX without images.

RayoMD 4.6.0

On 2026-10-14, the same machine, container and harness, only RayoMD measured again, the 4.6.0 and 4.4.0 release binaries in one run (median of 20): without images 1.37 ms on four cores and 1.39 ms on one (4.4.0: 1.27 and 1.30 ms in that run), with the PNG 7.64 and 11.24 ms (4.4.0: 7.58 and 11.06 ms); peak memory 3.6 MiB (4.4.0: 3.4 MiB); a further export in a running process 0.23 ms (4.4.0: 0.20 ms). 4.5 and 4.6 add work rather than remove it: formulas embed their fonts (the imageless PDF grows from 75 to 86 KB) and HTML is read. Against the other tools' figures above, that is 40× faster than glu, 51× than Qt 6, 252× than Pandoc + Typst and 1,530× than Pandoc + XeLaTeX without images.