#1313th of 22 in ML frameworks and libraries
jax
Python library for composable autodiff, JIT and vectorization of NumPy-style code
- Stars
- 36.4k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Sep 2026
- Language
- Python
34,988 on 25 Mar 202636,429 on 14 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
JAX is a Python array library that transforms NumPy-style functions with grad (forward and reverse-mode autodiff), jit (XLA compilation) and vmap (auto-vectorization). The transformations compose, and code runs on CPU, NVIDIA GPU, Google TPU and AMD GPU, with experimental Apple and Intel GPU support. Scaling across devices uses automatic, explicit or manual (shard_map) sharding.
Who it is for: ML researchers and engineers writing numerical or model code for accelerators
Strengths
- grad, jit and vmap compose arbitrarily, including higher-order derivatives and per-example gradients
- Differentiates through Python loops, branches, recursion and closures
- Three sharding modes (auto, explicit, manual) for scaling across many devices
- Compiles via XLA to CPU, NVIDIA, TPU and AMD targets using pip extras
Weaknesses
- README calls it a research project with sharp edges; not an official Google product
- jit restricts the Python control flow a function can use
- No native Windows GPU support; WSL2 GPU is experimental
- Apple and Intel GPU backends are experimental
What it needs
- GPU optional
- Needs XLA, NumPy
Also in ML frameworks and libraries
See all 22| Rank | Project | Score |
|---|---|---|
| 1 | transformersModel-definition library for text, vision, audio and multimodal models | 94 out of 100 |
| 2 | pytorchPython tensor library with GPU support and tape-based autograd for deep learning | 93 out of 100 |
| 3 | ultralyticsPython library and CLI for training and running YOLO vision models | 88 out of 100 |
| #4 | supervisionPython toolkit for loading, annotating and counting computer vision detections | 84 out of 100 |
| #5 | pytorch-image-modelsPyTorch library of image models, pretrained weights and reference training scripts | 82 out of 100 |