jax-js is a high-performance machine learning framework and compiler built specifically for the web, designed to bring JAX-style numerical computing to the browser. Developed from scratch with zero external dependencies, it provides an end-to-end environment for running neural networks, image algorithms, and complex simulations entirely client-side. The library focuses on portability, ensuring that developers can deploy compute-intensive applications that run efficiently across all major modern browsers including Chrome, Firefox, Safari, and their mobile counterparts.

The framework operates as both a library and a compiler, translating high-level array operations into an intermediate representation before synthesizing optimized kernels in WebAssembly (for CPU) and WebGPU (for GPU). By maintaining close API compatibility with NumPy and JAX, it offers a familiar development experience for those already accustomed to Python-based data science workflows. The library includes native support for functional transformations such as automatic differentiation, vectorization, and JIT compilation to fuse operations and maximize performance.

Some of the key features are:

• JAX-Style API: Maintains familiar API patterns similar to NumPy and JAX, making it intuitive for experienced data scientists.
• Cross-Backend Support: Includes backends for WebGPU (for high-performance hardware acceleration), WebAssembly (CPU), and WebGL2 for wider compatibility.
• JIT Compilation: Features a lightweight compiler that fuses mathematical operations to significantly improve memory bandwidth and execution speed.
• Advanced Autodiff: Provides native support for gradients, Jacobians, and Hessians, alongside forward-mode differentiation (jvp) and automatic vectorization (vmap).
• Zero Dependencies: Built from scratch in pure JavaScript, ensuring a lightweight footprint without external bloat.
• Memory Management: Utilizes an explicit reference-counting ownership model to manage memory efficiently in the JavaScript environment.
• Standardized Formats: Supports reading safetensors and provides modules for loading ONNX models and common optimizer implementations.

Users interact with jax-js through a modular JavaScript or TypeScript interface. When executing numerical code, the library tracks array operations and manages memory via a reference-counting system. By utilizing the jit() function, developers can group sequences of operations into single-kernel dispatches, bypassing the limitations of fragmented execution. The framework also allows for seamless switching between devices, enabling users to prioritize GPU acceleration when available or fallback to WebAssembly when necessary. All computations can be inspected and profiled directly within standard browser development tools.

Some common use cases include:

• In-Browser Neural Network Training: Implementing and training custom neural network architectures like those used for MNIST classification directly within a user's browser.
• Local AI Inference: Running small to medium language models (like Gemma 3) or voice-cloning models (like Kyutai Pocket TTS) locally to protect user privacy and remove server latency.
• Interactive Visualizations: Creating high-performance, interactive mathematical visualizations, such as fluid dynamics simulations or real-time Mandelbrot set rendering.
• Client-Side Data Processing: Leveraging WebGPU to compute embeddings or perform complex image transformations on user-uploaded datasets without transmitting sensitive information to external servers.