JAX is a free, open-source Python library from Google for high-performance numerical computing and machine-learning research. It has earned more than 36,000 stars on GitHub and is under very active development, with its latest updates in 2026. JAX combines a familiar NumPy-style API with automatic differentiation and just-in-time compilation to GPUs and TPUs.
The core idea is composable function transformations: grad differentiates your code, jit compiles it with XLA for accelerators, vmap vectorizes it over batches automatically, and pmap spreads it across devices — and all of them compose freely with each other. You write ordinary NumPy-like Python, and JAX turns it into optimized accelerator code. This functional, transformation-first design has made JAX the foundation of a large research ecosystem, including Flax for neural networks, Optax for optimization and the training stacks behind many well-known models. Large-scale training frameworks for language models, robotics and scientific computing are likewise built directly on these primitives. Its strict functional style — pure functions, explicit random-number handling and no hidden state — makes experiments unusually reproducible, a property researchers prize when results have to be trusted and replicated.
JAX is completely free and open source under the Apache-2.0 license. There is no paid tier; your only cost is the compute you run it on, from a laptop CPU to a TPU pod.
JAX is for ML researchers and engineers who want NumPy’s ergonomics with accelerator performance and first-class differentiation — especially those training large models or doing research that pushes beyond standard frameworks.