JAX
docs.jax.dev
- Free plan
- Windows
- macOS
- Linux
- Self-hosted
What JAX is
JAX is a free, open-source Python library for accelerator-oriented array computation and program transformation. It targets high-performance numerical computing and large-scale machine learning, with a NumPy-style API for researchers and engineers. Composable transformations support compilation, batching, automatic differentiation and parallelization. The same code can run on CPU, GPU or TPU backends, with GPU and distributed training supported. JAX’s core remains focused on efficient array operations and transformations, while an evolving ecosystem includes Flax, Keras, Optax, TensorFlow Datasets, Hugging Face Datasets, NumPyro and PyMC. Installation uses the `jax` package together with platform- and accelerator-specific `jaxlib` binaries. It supports local training and runs on Linux, macOS, Windows and self-hosted environments. The project is developed openly on GitHub and welcomes contributions. Windows x86_64 CPU and Apple GPU support on Apple ARM are marked experimental.
Who it suits
JAX suits Python researchers, engineers and machine-learning teams working with numerical computing at scale. It is also aimed at users who need CPU, GPU or TPU execution and distributed training.
Good to know, and what to watch
- NumPy-style API for Python users
- Composable compilation and autodiff transformations
- CPU, GPU and TPU backends
- Supports distributed training
- Free open-source library
- Windows x86_64 CPU support is experimental
- Apple ARM GPU support is experimental
- Distributed connections are plaintext by default
- JIT results can differ numerically from eager runs
Plans
Every plan and priceJAX
Free plan
The verdict
JAX provides a focused foundation for large-scale numerical computing and machine learning, with transformations and multiple accelerator backends. Users should review its experimental platform support, security defaults and possible JIT numerical differences before deployment.
What the maker publishes
- Accelerator support
- The project describes its core library as focused on machine learning and numerical computing, with a modular backend stack for targeting different accelerators. source
- Contributors
- The project says it welcomes open-source contributions and frequently receives contributions from Google DeepMind, Alphabet, NVIDIA and others. source
- Development
- JAX development takes place in the open on GitHub through pull requests, the issue tracker, discussions and JAX Enhancement Proposals. source
- Ecosystem
- The documentation lists ecosystem tools including Flax, Keras, Optax, TensorFlow Datasets, Hugging Face Datasets and NumPyro. source
- Hardware
- The same code can execute on multiple backends, including CPU, GPU, and TPU. source
- Installation
- JAX uses the `jax` Python package and `jaxlib`, which contains compiled binaries with different builds for operating systems and accelerators. source
- Integrations
- The documentation lists ecosystem libraries including Flax, Keras, Optax, TensorFlow Datasets, Hugging Face Datasets, NumPyro, and PyMC. source
- Numerical behavior
- The FAQ explains that XLA optimizations can cause JIT-compiled computations to produce numerically different results from eager computations. source
- NumPy-style API
- JAX provides a familiar NumPy-style API for researchers and engineers. source
- Open development
- JAX development takes place in the open on GitHub, where the project uses pull requests, an issue tracker, discussions, and JAX Enhancement Proposals. source
- Platform limits
- The installation guide labels Windows x86_64 CPU support experimental and Apple GPU support on Apple ARM experimental. source
- Platforms
- The installation guide lists CPU support for Linux, macOS and Windows, with accelerator support varying by operating system and hardware. source
- Scope
- The JAX project says the core library is narrowly scoped around efficient array operations and program transformations, with an evolving surrounding ecosystem. source
- Security
- Distributed coordination connections are neither encrypted nor authenticated by default; the documentation describes optional mutual TLS for securing them. source
- Security limits
- The documentation says CPU and non-NVLink GPU collectives, and the JAX profiler server connection, are plaintext and unauthenticated. source
- Support
- The installation guide directs users with problems using prebuilt wheels to the JAX GitHub issue tracker. source
- Transformations
- JAX includes composable transformations for compilation, batching, automatic differentiation, and parallelization. source
- What it does
- JAX is a Python library for accelerator-oriented array computation and program transformation, designed for high-performance numerical computing and large-scale machine learning. source
- Who it is for
- The project describes its core library as focused on the fundamentals of machine learning and numerical computing at scale. source
Checked against the maker's pages on 11 Oct 2026.