Keras is a high-level Python library for making, training and deploying deep learning models, which runs on top of JAX, TensorFlow or PyTorch.
In plain words
Deep learning frameworks can be verbose. Keras puts a friendly design first so you can build quickly: you stack layers into a model, then call fit to train it. Since Keras 3, the same code can run on JAX, TensorFlow or PyTorch, so you can switch to whichever is fastest for your model. KerasHub adds ready-made versions of popular model designs with pretrained weights.
A quick example
A typical first model: stack a few layers to classify handwritten digits from MNIST, call fit to train it, then evaluate it on test images. Changing one setting runs the same code on a different backend.
Where you'll meet it
- Learning deep learning and quick prototyping.
- Research and production at scientific organisations.
Related
Sources
- DocsKeras: The multi-backend deep learning library, Keras · read 5 Nov 2026
- DocsIntroducing Keras 3.0, Keras · read 5 Nov 2026
Checked by WhatIsAI against the sources above · 5 Nov 2026. Spotted an error? Report it.