Keras 3's Breakthrough

One API, three engines

One API, three backends: TensorFlow, PyTorch, and JAX (plus OpenVINO for inference-only). The latest stable version is Keras 3.11.3.

The breakthrough is the keras.ops namespace — a universal operation layer that implements the full NumPy API plus neural network-specific functions. When you write keras.ops.matmul(x, w), Keras dispatches the call to jax.numpy.matmul, tf.matmul, or torch.matmul depending on the active backend, producing numerically identical results.

What "backend-agnostic" buys you

Write your model once, train it on any framework
Save a model trained on JAX, load it in TensorFlow for deployment
Use PyTorch's ecosystem for research, then export for TF Serving in production
Every built-in Keras layer, metric, loss, and optimizer uses keras.ops internally — so they're all portable for free

That last point is the quiet one that matters most: portability isn't a feature you opt into, it's the default the whole library is built on. The Code block below is the layer that makes it true.

# keras.ops — the universal operation layer import keras # These work identically on TF, PyTorch, or JAX: x = keras.ops.matmul(a, b) y = keras.ops.nn.softmax(logits) z = keras.ops.image.resize(img, size=(224, 224))

One API, three engines

What "backend-agnostic" buys you

Code

External links

Exercise

Progress

Comments 0