JAX
Google's Python library for accelerator-oriented array computation: jit, grad, vmap and sharding across CPU, GPU and TPU
Overview
JAX is "a Python library for accelerator-oriented array computation and program transformation." You write NumPy-style code and compose transformations on it: jax.jit compiles it with XLA, jax.grad differentiates it, jax.vmap vectorizes it, and sharding and shard_map spread it across many devices. The same code runs on CPU, GPU and TPU. Neural networks, optimizers and checkpoints come from companion libraries (Flax, Optax, Orbax). Google's 2026 messaging leans on it: the TensorFlow 2.21 post recommends Keras 3, JAX and PyTorch for new generative AI work, and a May 2026 Google Cloud post says "the future of scalable machine learning is being built on JAX." JAX is Apache-2.0 licensed; v0.12.0 (October 19, 2026) now needs NumPy 2.2+ and a Haswell or newer x86-64 CPU. Also from Google: TensorFlow, Keras (which runs on JAX), Gemma and Colab.
The Verdict
Who Should Use JAX?
Best For
- Large-scale training on TPUs or many GPUs
- Researchers who want composable autodiff and vectorization
- Teams migrating TensorFlow models onto Google's newer stack
Not Ideal For
- Mac users who need GPU acceleration (macOS is CPU only)
- Windows users on NVIDIA GPUs (native support not available; WSL2 experimental)
- Developers who want a batteries-included, object-oriented API out of the box
What's Great
- Composable transformations: jit, grad, vmap and sharding work together
- Same code on CPU, GPU and TPU, with automatic parallelization
- Keras 3 says JAX "typically delivers the best training and inference performance" across its backends
- Frequent releases (monthly through 2026)
Watch Out For
- Platform gaps: no Mac GPU, no native Windows NVIDIA, Intel GPU experimental
- Breaking changes in minor releases (v0.12.0 dropped NumPy 2.1 and pre-Haswell CPUs)
- Functional, stateless style is a shift from TensorFlow's stateful layers
Pricing
View all features & details
Key Features
- jax.jit with ahead-of-time compilation
- jax.grad, JVP, VJP, Jacobians and Hessians
- jax.vmap automatic vectorization
- Sharding and shard_map for multi-device parallelism
- v0.12.0: custom_vjp.defremat, mxfp4 support, IFRT Proxy backend for Pathways
Platforms
- Linux x86_64: CPU, NVIDIA GPU (CUDA 13), Cloud TPU, AMD GPU
- Linux aarch64: CPU and NVIDIA GPU
- macOS: CPU only
- Windows: CPU experimental, WSL2 experimental
Real-World Usage
At Google
- Google Cloud: "For many teams at Google and across the industry, the future of scalable machine learning is being built on JAX"
- Google publishes tooling to migrate TensorFlow production models to JAX
How It Compares
| Feature | JAX | PyTorch | TensorFlow | MLX |
|---|---|---|---|---|
| Programming model | Functional transforms | Eager, object-oriented | Graphs plus Keras | NumPy-like, lazy |
| TPU | Native | TorchTPU backend | Not covered here | No |
| Keras 3 backend | Yes | Yes | Yes | Named on keras.io |
| Mac GPU | No | MPS | Not covered here | Yes, Apple silicon first |
| License | Apache-2.0 | BSD-style | Apache-2.0 | MIT |
Sources: GitHub · JAX install docs · PyTorch Conference 2026 · MLX docs
Is this your tool? Add a badge