JAX neural nets you can actually mutate like Python objects

What it does

Flax is the dominant high-level neural network ecosystem for JAX, maintained by a dedicated team at Google DeepMind. It provides standard layers like Linear, Conv, and MultiHeadAttention, plus utilities for checkpointing, replicated training, and device prefetching. The project recently introduced Flax NNX, a simplified API that lets you define models with standard Python classes and mutation rather than wrestling with pure functional transformations.

The interesting bit

The big shift is philosophical: the original Linen API (2020) leaned heavily into JAX’s functional roots, but NNX explicitly reintroduces reference sharing and mutability. That means you can debug a model by simply inspecting a Python object or share state between layers without bending the framework into a pretzel.

Key highlights

• First-class support for Python reference semantics, allowing standard object mutation and state sharing.
• Ships with standard layers (Linear, Conv, BatchNorm, LSTMCell, GRUCell, MultiHeadAttention) and utilities for checkpointing, metrics, and replicated training.
• Maintained by a Google DeepMind team in close collaboration with the JAX developers.
• Explicitly built for research flexibility: the intended workflow is to fork an example and modify the training loop, not bolt features onto a monolithic framework.
• Includes educational examples from MNIST to Gemma transformer inference.

Verdict

If you are doing research in JAX and want a library that gets out of your way when you need to hack on a training loop, this is the default choice. If you are committed to PyTorch or Keras and have no interest in JAX, this will not change your mind.