Compared with similar libraries

flowdown next to five other React Markdown renderers: Streamdown 2.7, react-markdown 10, markdown-to-jsx 9.10, markstream-react 2.0 and marked-react 4. Each one is set up the way its own README sets up a chat, without plugins, and gets the same replies.

Every figure and mark here comes from a script in bench that you can rerun. The few rows taken from a library's documentation say so. The sizes, package counts and feature marks were rerun and checked with Claude Fable 5.2 on 2026-11-10 (what was checked).

While a reply streams#

What the reader sees in a frame where the reply stops mid-syntax.

¹ It waits for the closing **, so the word shows unbolded until then.

Markdown#

All six draw tables, task lists, strikethrough and bare links.

All six drop onerror handlers and <script>, render on the server with renderToString, and leave no javascript: link that runs. marked-react keeps the link, but React 19 swaps its address for one that only throws.

How these are checked: the same inputs go through every renderer, and each mark is a check on the DOM it drew (bench/src/features.tsx). The markup behind every mark is in bench/results/features.json.

In your project#

• Download is what loads with the component. Each library is bundled as a page's bundler would do it: a production build, minified and gzipped, without React; react-markdown with remark-gfm (bench/src/size.ts).
• Loaded on demand. flowdown fetches its HTML reader and the full table of HTML5 character references (14 KB together) only for a reply that needs them. Streamdown loads its HTML parser up front. Packages a renderer loads only when a reply needs them (KaTeX, Mermaid, D2) are not counted for any library.
• Highlighting. flowdown with code highlighting for 23 languages is 68 KB.
• Packages installed counts dependencies all the way down, read from the lockfile, without React (bench/src/install.ts).

Code, math and diagrams#

From each library's documentation, not measured.

Speed while streaming#

Main-thread time to stream a whole reply, one update per animation frame, in milliseconds. Lower is better.

• Where flowdown wins: everywhere here, and most on length. It parses again only the block still being written, and reads only the new lines of a growing code block, so a 100k reply and a 30-reply thread cost about as much per update as the first lines.
• Where the others come close: short replies and one long code block. markdown-to-jsx is within 5 ms on a 5k chat answer and within 2× on 50k of code, where marked-react takes a little over 2×. markdown-to-jsx holds unfinished syntax back instead of repairing it.
• Why react-markdown is slowest. It parses the whole text again on every update.

How speed is measured: the whole table three times over on an Apple M5 in Bun 1.4.4, three runs of each stream every time, and the median kept. Pages are drawn with happy-dom, not a real browser, so the figures compare renderers rather than predict a real page. Every scenario, chunk pattern and percentile is in bench/results/latest.md.

Parser speed#

The parser alone, with no React and no DOM: megabytes of Markdown read in a second, each text of the benchmark parsed whole. Higher is better.

• What is compared. flowdown's parser is the one <Markdown> runs. mdast-util-from-markdown is remark's, the one react-markdown and Streamdown run; micromark is the tokenizer under it. All read GitHub's extensions: tables, task lists, strikethrough, autolinks and, except marked, footnotes.
• A tree against HTML. marked and markdown-it write HTML and build no tree. flowdown writing HTML from its tree reads 131 and 139 MB in a second.
• Where the trees differ. mdast gives every node its position in the source; flowdown's tree gives positions to blocks only, which is part of the gap.
• In a reply's terms. A 100k-character reply is parsed whole in under 0.6 ms, so the parser is a small share of what streaming one costs: the rest is React and the DOM.

How it is measured: the median of three runs on an Apple M5, each parser in a process of its own, in Bun 1.4.4 and Node 25.6.1. Every text, and what each parser read in it, is in bench/results/parse.md.

Which one to pick#

• flowdown. For an AI chat or assistant whose replies run long, threads that grow, and documentation. It repairs unfinished syntax, comes styled, and is one package of 31.7 KB with its own parser. It does not take remark or rehype plugins.
• markdown-to-jsx. When a small, unstyled renderer for short replies is enough. It is one package and 31 KB. It slows down on very long replies and threads, and ships no styles.
• Streamdown. If the app is already on Tailwind and shadcn/ui. It comes from Vercel, with plugins for Shiki, math, Mermaid and CJK.
• react-markdown. For finished Markdown and the remark and rehype plugins, not for streaming.
• markstream. For D2 diagrams and code diffs as a reply streams. It is the largest download.
• marked-react. For small, finished Markdown. It has no streaming repairs, footnotes or raw HTML.

Run it yourself#

From bench/:

bun run size      # download sizes
bun run features  # what each renderer draws
bun run deps      # packages installed, licence, styles
bun run bench     # streaming speed, about half an hour
bun run charts    # redraw the README chart from the results

flowdown's answer forms (steps, tabs, stats, pros and cons, folded sections, sources) are its own dialect, so they are not compared here.