jax-js-nonconsuming (fork) is a machine learningan ML library and compiler for the web
Fork Notice
This is a non-consuming ownership fork of ekzhang/jax-js. Operations leave inputs alive (no manual .ref needed), designed
for teams familiar with NumPy or MATLAB.
High-performance WebGPU and WebAssembly kernels in JavaScript. Run neural networks, image algorithms, simulations, and numerical code, all JIT compiled in your browser.
Add jax-js-nonconsuming to your project
Zero dependencies. All major browsers. Install from or .
pnpm users: add "pnpm": { "onlyBuiltDependencies": ["@hamk-uas/jax-js-nonconsuming"] } to your package.json so pnpm allows the Git dependency's prepare build.
Matrix multiplication
Billions of floating-point operations (GFLOPs) per second
Running benchmark…
Try it out!
This is a live editor, the code is running in your browser.
Run code to see output here.
Learn more
GitHub Repository
Get started with jax-js-nonconsuming and check out the tutorial.
REPL
Try out the library in this browser-based REPL.
API Reference
View the full API documentation.
MobileCLIP2 Inference
Generate embeddings for books and search them in real time.
Kyutai Pocket TTS
Voice cloning AI model that runs in your browser.
MNIST Training
Demo of training a neural network on MNIST.
DETR ResNet-50
Object detection with DETR via ONNX, running in-browser.
Mandelbrot Set
GPU-accelerated fractal rendering with JIT-compiled scan.
Fluid Simulation
Real-time Navier-Stokes fluid simulation on WebGPU.
Matmul Benchmark
Measure matrix multiplication GFLOP/s on your device.
Tiled Matmul Sweep
Run all tile configs on your GPU and compare GFLOP/s.
Conv2d Benchmark
Measure convolution throughput on your device.