Zyth
WebAssembly · WebGPU · React · no framework at all

Your model,
running in the page.

No server to rent for inference. A Zyth program builds into a web page or an npm package, and its tensors and models run in the visitor's browser, on their GPU when there is one.

1 · a page

The whole program in a page

Everything Zyth does, in the browser: tensors, training, GPU drawing.

zyth build --target web app.zyth -o dist
2 · a component

A React component in npm

Props checked, TypeScript types generated, one shared runtime for every Zyth package in the app.

zyth build --target npm-jsx xor.zyth -o xor
npm install ./xor @biodream-llc/zyth-runtime
3 · a link

Shared at its own address

Publish a web build and send the link. Anyone can open it, and remix its source.

zyth share app.zyth
# https://<name>.zyth.io
App.tsx
import { XorTrainer } from "xor";
import { Cube } from "cube";

export function App() {
  return (
    <main>
      <XorTrainer steps={50} />
      <Cube />
    </main>
  );
}

It is just React

A Zyth component is a function that returns elements, with the hooks you know: use_state, use_effect, use_reducer, use_context, use_memo, and Redux's use_selector and use_dispatch. Every HTML and SVG tag is there.

The difference is what it can hold: a neural network that trains when the visitor clicks, a language model that answers questions in the tab (talk to it), a 3D scene drawn on the GPU (see the solar system).

The runtime package is on npm today.

Or no framework at all

zyth.web.ui is Zyth's own way to build pages. A component builds its part of the page once, then changes only the nodes whose values changed: no virtual DOM, nothing drawn twice, no React in the bundle. State, derived values, keyed lists, events and routing are built in, and resource waits for the WebAssembly half when a page calls a model.

The JavaScript a page downloads is 2.7 KB gzipped. The same table in Svelte 5 is 13.8 KB, in React 19 68.4 KB.

counter.zyth
use zyth.web.ui.{div, h1, button, state, View}

@export
fn Counter(label: str, start: i64 = 0) -> View {
    var count = state(start)
    return div(class: "counter") {
        h1 { "{label}: {count}" }
        button(on_click: || { count += 1 }) { "+1" }
    }
}

As fast as Svelte, measured

The operations of js-framework-benchmark on a table of rows, milliseconds from the click to the frame that shows the result, median of 30 runs, production builds, headless Chromium. Svelte 5 written for speed, React 19 with memo.

OperationZythSvelte 5React 19
create 1,000 rows44.446.155.6
replace 1,000 rows52.054.268.0
update every 10th row14.614.718.1
select a row2.12.73.7
swap two rows5.35.344.7
remove a row7.38.110.5
create 10,000 rows428.8430.2819.6
append 1,000 rows51.251.761.8
clear 1,000 rows4.96.59.4

Level with Svelte on most operations, ahead on select, remove and clear, and ahead of React on all of them. Not measured yet: memory, start-up, Firefox and Safari.

A smaller runtime for pages

Pages that don't draw or decode images get Zyth's light WebAssembly runtime: 3.5 MB, 1.2 MB gzipped, with programs already checked at build time. It loads in 0.64 to 0.69 of the full runtime's time at the same speed: 75 ms against 115 from a local server, 1.24 s against 2.10 s at 10 Mbit/s.

What runs in the demos

  • A tiny GPT trained on Shakespeare, writing text in the page.
  • A network learning XOR inside a React + Redux app, with a 3D cube drawn by Zyth beside it.
  • A shape error caught in a live editor, before anything runs.
  • Language models on the visitor's GPU: a chat assistant running Qwen2.5-Instruct through WebGPU, about 120 tokens a second on an RTX 3090 with the 0.5B model. The build packs the weights to 4 bits and the page keeps them, so the second visit starts without a download.
  • GPU drawing through WebGPU: the solar system, with 24,000 asteroids moved as GPU tensors every frame and a glow pass. A million asteroids run at 519 frames a second on an RTX 3090, against 19 for three.js moving them from JavaScript.

Open the demos