JAX is a Python library for high-performance numerical computing and large-scale machine learning, with a focus on accelerator-oriented array computation and program transformation. Developed by the JAX authors, it offers a familiar NumPy-style API designed to help researchers and engineers adopt its capabilities. JAX provides composable transformations for just-in-time compilation, automatic differentiation, batching, and parallelization, and the same code can run across CPU, GPU, and TPU backends.

The project provides installation guidance, tutorials, API references, and advanced documentation for working with arrays, random numbers, gradients, and program state. Its learning resources progress from expressing computations to performance and scaling, advanced autodiff, custom kernels, distributed systems, and JAX internals. Users can also explore tools for profiling, debugging, sharding, serialization, and connecting external code through a foreign function interface. JAX itself is deliberately focused on efficient array operations and transformations, while a broader ecosystem supplies complementary tools for neural networks, optimization, data loading, probabilistic programming, simulation, and language models.