Skip to main content

Ridiculously fast Markdown for React and AI.

Open playgroundRead the docsNPM
$npm install @aeven-ai/hypermarkdown
1.6×–10.6×Faster than the nearest streaming renderer across the suite.
659 msvs 4,417 ms on a captured AI code stream.
654 msCaptured AI table. Streamdown: 11,621 ms.

Production streaming benchmark

Apple M2 Max, Node v26.7.0. Chunk processing plus synchronous React commit. Same Markdown. Same stream.

Methodology →
WorkloadHyperMarkdownMarkstreamStreamdownDeepSeek Harnessreact-markdown
Large code block216 ms769 ms3,242 ms4,416 ms2,054 ms
Mixed prose153 ms313 ms559 ms249 ms2,629 ms
Captured AI code659 ms4,417 ms12,511 ms14,621 ms8,008 ms
Captured AI table654 ms4,948 ms11,621 ms19,644 ms10,236 ms
Large table874 ms9,276 ms33,142 ms55,918 ms29,747 ms

Completed work stays completed

Block-level freezing cannot help a 1,000-line fence that is still open. HyperMarkdown caches inside the active block.

Typical stream renderer

  • new token
  • growing active block
  • parse the block again
  • render again

HyperMarkdown

  • settled code lines → cached
  • settled table rows → cached
  • settled list items → cached
  • changing frontier → parse
new token
  1. const line = alreadySettled;cached line
  2. | already | settled |cached row
  3. - already a list itemcached item
  4. const frontier = stillGrowingparse

Installation

React 18 or 19 as a peer. Heavy engines stay optional.

npm install @aeven-ai/hypermarkdown
import "@aeven-ai/hypermarkdown/styles.css";
import { HyperMarkdown } from "@aeven-ai/hypermarkdown";

<HyperMarkdown md={markdown} />

Streaming example

Mount one renderer. Write each delta. Finalize once.

const ref = useRef<HyperMarkdownHandle>(null);

for await (const delta of stream) {
  ref.current?.write(delta);
}
ref.current?.write("", true);

<HyperMarkdown ref={ref} streaming />

Sub-block cache

Settled code lines, table rows, and list items are not parsed again.

Incomplete Markdown

Half-written links, tags, and math are withheld until they are safe.

GFM built in

Tables, task lists, autolinks, strikethrough, and footnotes.

Optional plugins

KaTeX, highlight.js, Mermaid, and CJK load only when you ask.

Reasoning blocks

<think>, <thinking>, and <reasoning> stream into a collapsible panel.

React 18 and 19

Finished Markdown can render on the server; the handle stays on the client.

Used by Æven

HyperMarkdown is the official Markdown component used by Æven and integrates with the DeepSeek Harness (DSH) architecture. It was built around long answers, dense code, wide tables, reasoning traces, and many small deltas — not adapted from a finished-document renderer after the fact. The captured AI workloads in the benchmark suite come from that environment.