jax-ml/jax

Health v2 · maintenance only

Adoption not part of the score

36215 stars · 3744 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

JAX is a Python library for accelerator-oriented array computation and composable program transformations, offering NumPy-compatible APIs with automatic differentiation, vectorization, and JIT compilation via XLA. It scales the same code across CPU, GPU, and TPU backends and includes Pallas for writing custom hardware kernels.

Use cases

• train neural networks with automatic differentiation in Python
• JIT-compile NumPy-style code to run on GPUs and TPUs
• vectorize per-example gradient computations with vmap
• write custom GPU/TPU kernels with Pallas
• run large-scale machine learning research across accelerators
• compute higher-order derivatives like Jacobians and Hessians
• parallelize and shard computations across multiple devices

When to choose

• you need high-performance numerical computing or ML research on GPU/TPU
• you want composable autodiff, batching, and compilation in pure Python
• you prefer a NumPy-like API with accelerator scaling
• you need custom low-level kernels for TPU or Hopper+ GPUs

When to avoid

• you need a batteries-included deep learning framework with prebuilt layers and training loops out of the box
• your project depends on impure Python code with side effects inside compiled functions
• you need stable experimental APIs like Pallas Triton or unsupported hardware backends
• you only need simple CPU array math where plain NumPy suffices

Facets

library · maturity stable

machine-learning deep-learning math compiler gpu-computing llm-training machine-learning deep-learning gpu-computing data-science python windows cloud autodiff jit-compilation xla numpy tpu vectorization pallas numerical-computing algorithms linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="jax-ml/jax")