JAX is an open-source Python library for array computing, the kind of numerical work done with matrices and vectors in science and machine learning. Its code is published on GitHub by the jax-ml project and copyrighted by "The JAX Authors"; the README calls it a research project, not an official Google product.
A developer writes ordinary Python functions with jax.numpy, an API that copies NumPy. JAX then applies function transformations to them: jax.grad performs automatic differentiation (it computes exact derivatives of the code, in reverse mode, also called backpropagation, or forward mode, and through loops and branches); jax.jit compiles a function with XLA, a compiler for linear algebra, into fast code for the hardware; jax.vmap turns a function for one example into one for a batch; and sharding tools split arrays across many devices. The same code runs on CPUs, NVIDIA and AMD GPUs and Google TPUs. Other libraries, such as Flax and Optax for neural networks, are built on it.
JAX is free and open source under the Apache 2.0 license and installs with pip.