JAX is a Python library for numerical computing developed by Google Research and released as open source in 2018. It offers an array API nearly identical to NumPy while adding function transformations for automatic differentiation (grad), just-in-time compilation (jit), automatic vectorization (vmap), and multi-device parallelism (pmap). Computations are compiled through XLA and run on CPUs, GPUs, and TPUs.
Where PyTorch centers on object-oriented model definitions, JAX is designed around pure functions and their transformations, which has made it popular for research code and large-scale TPU training. Google DeepMind and other research organizations use it to train large models, and higher-level libraries such as Flax and Haiku provide neural network layers on top of it.