jax-js is a machine learning framework designed for use in web browsers, offering JAX-style capabilities in JavaScript. It allows developers to run high-performance numerical applications client-side, leveraging WebAssembly and WebGPU for efficient computations. The library aims for compatibility with NumPy and JAX APIs, making it accessible for those familiar with those tools. With zero external dependencies, jax-js is lightweight, with a bundle size of just 80 KB, significantly smaller than alternatives like TensorFlow.js (269 KB) and onnxruntime-web. The framework supports various advanced features such as automatic differentiation and JIT compilation, which optimizes performance by fusing operations into single kernel dispatches. It enables users to perform complex operations like matrix multiplication, general einsum, and basic convolutions efficiently. The reference-counted ownership model in jax-js manages memory automatically, reducing the risk of memory leaks during computations. This approach requires users to be mindful of their reference counts but ultimately leads to better memory management. Community contributions demonstrate the framework's versatility, with projects ranging from training neural networks on standard datasets like MNIST to voice cloning and in-browser object detection. Users can easily get started by importing the library through a module script tag in HTML. This simplicity, combined with the framework's robust features, positions jax-js as a practical option for web-based machine learning tasks.