TypeScript icon, indicating that this package has built-in type declarations

0.0.4 • Public • Published


A NumPy-like library in the browser, running on the GPU, with automatic differentiation and JIT compilation.

Accelerated numerical computing extensions for JavaScript.

This is a numerical computing library that runs in the browser through WebGPU, enabling fast matrix and linear algebra operations for machine learning, computer graphics, generative art, and scientific computing.

Getting Started

Install the package to add it to your JavaScript project.

npm install nqx

Then, you can start writing code:

import { createEngine } from nqx;

const nqx = createEngine("webgpu");
await nqx.ready();

const x = nqx.array([1, 2, 3]);
const y = nqx.array([4, 5, 6]);
const z = x.add(y);

await z.js();                   // [5, 7, 9]
await z.print();                // array([5., 7., 9.], dtype=float32)
await x.mul(y.add(1)).print();  // array([5., 12., 21.], dtype=float32)

Non-Bundler Usage

If you don't want to use a bundler or don't have an appropriate toolchain, you can import nqx directly from a CDN.

<script type="module">
  import { createEngine } from "https://unpkg.com/nqx/dist/nqx.js";

  const nqx = createEngine("webgpu");
  await nqx.ready();

  const x = nqx.array([1, 2, 3]);
  await x.print();

If you're not using JavaScript modules, you can still use a script tag. This allows you to access the library from the NQX global variable.

<script src="https://unpkg.com/nqx/dist/nqx.iife.js"></script>
  (async () => {
    const nqx = NQX.createEngine("webgpu");
    await nqx.ready();

    const x = nqx.array([1, 2, 3]);
    await x.print();


This library is written in TypeScript and developed primarily using Vite. Run npm install to download dependencies and npm test to run tests, fully in the browser. You may need to install Chrome web driver components as prompted.

To bundle the library for distribution, run npm run build. This produces build artifacts in the dist folder.


This library is inspired by influential work in the Python ecosystem: NumPy and JAX, as well as the related libraries derived from them (to varying degrees: PyTorch, Tensorflow, Xarray, Dask, etc.). It's meant to fill the void of a simple, fast, and reliable system for numerical coding. What would be the JavaScript or web-based equivalent of that?

The state-of-the-art for GPU-accelerated computing in JavaScript has developers writing their own WebGL or WebGPU shaders, which is awkward and time-consuming. No Python developer would ever write their own CUDA kernels or BLAS implementations!

Compared to TensorFlow.js, nqx is much simpler and smaller, and it provides some key advantages:

  • Uses native JavaScript data types and has a concise API.
  • Not limited to machine learning use cases.
  • More flexible, offering forward and reverse-mode automatic differentiation for arbitrary functions, including second and higher derivatives.
  • (Planned) Includes a tracing JIT compiler for numerical operations with automatic kernel fusion.
  • (Planned) Has a multithreaded WebAssembly backend with clean exchange between arrays stored on CPU and GPU devices.

However, there are some drawbacks:

  • Newer library, not backed by a team of paid developers at Google.
  • No support for binary Node.js and legacy WebGL backends.
  • No plans for adding a Keras-like API.


nqx is provided under the MIT license. See LICENSE.

Package Sidebar


npm i nqx

Weekly Downloads






Unpacked Size

207 kB

Total Files


Last publish


  • ekzhang