Flax

For researchers already committed to JAX, Flax supplies the model-building layer that JAX itself leaves out. Its current API, NNX, lets you write a network as an ordinary Python class whose attributes hold parameters and state, then train it with lifted transforms such as nnx.jit and nnx.grad that understand those mutable objects. The older Linen API, which keeps parameters in separate immutable trees, still ships and is documented separately, so existing codebases are not stranded. Out of the box you get the usual layers (dense, convolution, normalization, attention, dropout), an nnx.Optimizer wrapper around Optax, checkpointing through Orbax, and worked examples ranging from MNIST to Gemma sampling. Installation is pip install flax on top of a JAX build matched to your CPU, NVIDIA GPU or Cloud TPU, and a Google DeepMind team ships point releases every month or two. Apache 2.0 licensing permits commercial use. It suits ML researchers and teams training custom architectures, especially on TPUs, while newcomers who simply want to run existing models will find PyTorch-based tools easier to start with.