Keras 3: the multi-backend deep learning framework

What is Keras 3?

Keras 3 is a multi-backend deep learning framework that runs the same Python API on top of JAX, TensorFlow, or PyTorch, with an inference-only OpenVINO backend for deployment. Earlier Keras versions shipped bundled inside TensorFlow as tf.keras, tied to that one engine; Keras 3 splits the API from the execution engine so the backend is a choice per project, not a fixed dependency.

Key features of Keras 3

• ✓Runs unmodified on JAX, TensorFlow, or PyTorch — set KERAS_BACKEND (env var or ~/.keras/keras.json) before importing keras, and the same model code executes on whichever engine you picked.
• ✓Custom layers, models, and metrics written against the Keras API can drop into a raw training loop written in native TF, JAX, or PyTorch.
• ✓Accepts tf.data.Dataset pipelines or PyTorch DataLoaders as input regardless of which backend is active.
• ✓Ships a .keras model save format the README positions as the successor to the older TensorFlow SavedModel path for tf.keras users.
• ✓OpenVINO backend for inference-only deployment, separate from the three training-capable backends.
• ✓Existing tf.keras models without custom components run on JAX or PyTorch with no code changes; ones with a custom train_step or custom layers usually need a small backend-agnostic rewrite.

Installing Keras 3 and choosing a backend

Install Keras 3 from PyPI with `pip install keras --upgrade`, then add one backend package on top: `tensorflow`, `jax`, or `torch`. The OpenVINO backend installs separately and only covers inference, not training. Keras 2 is still available, packaged separately as `tf-keras`, if you're not ready to move. Keras 3 supports Linux and macOS directly; on Windows, the README points to WSL2. Minimum backend versions per the compatibility table: TensorFlow 2.16.1, JAX 0.4.20, PyTorch 2.1.0, OpenVINO 2026.2.0. Building from source: install `requirements.txt`, run `python pip_build.py --install` from the repo root, and run `./shell/api_gen.sh` if your PR touches the `keras_export` public APIs. For GPU support, the plain `requirements.txt` only pulls CPU wheels — you need the backend-specific `requirements-{backend}-cuda.txt` file, and the README recommends a separate environment per backend (its example uses `conda create -n keras-jax python=3.10`) to avoid CUDA version conflicts between backends.

Writing backend-agnostic code with Keras 3

Set the backend before you import anything: export `KERAS_BACKEND=jax` (or `tensorflow`/`torch`) in your shell, or edit the `backend` key in `~/.keras/keras.json`. In a notebook like Colab, set `os.environ["KERAS_BACKEND"]` before the `import keras` line. Once imported, the backend is locked for that process. If you're moving an existing tf.keras project, the README calls Keras 3 a drop-in replacement on the TensorFlow backend: switch `model.save()` calls to the `.keras` format and the rest should work unchanged. A model with no custom layers or custom train_step often runs on JAX or PyTorch immediately with zero code changes; one with custom components usually needs a short rewrite to a backend-agnostic form, which the README describes as normally taking a few minutes rather than a rearchitecture. Data loading isn't backend-locked either — the same model can consume a tf.data.Dataset pipeline or a PyTorch DataLoader, regardless of which of the three training backends is active.

Alternatives to Keras 3

Frequently asked questions

The problem it solves

A model built with tf.keras only runs on TensorFlow's engine, and a model built with native PyTorch or Flax/JAX only runs on that engine — so a team that picks TensorFlow early and later wants JAX's training speed, or vice versa, ends up rewriting the model-building code, not just the training script. Keras 3 exists to decouple the layer and model API from the execution engine so that rewrite isn't necessary.

Best use cases

• •Migrating an existing tf.keras project off TensorFlow lock-in without a full rewrite.
• •Prototyping on PyTorch's eager execution, then moving the same model to JAX for a faster training run.
• •Building a single codebase for vision, NLP, audio, or timeseries models that still has to run on backend-specific infrastructure.
• •Deploying trained models for inference only via the OpenVINO backend.

Who should try it — and who should skip

Try Keras 3 if you already write tf.keras-style code and want the option to train on JAX or PyTorch without maintaining three separate model definitions, or if you're picking a first deep learning framework and want to defer the TensorFlow-vs-PyTorch decision. Skip it if your team has deep, framework-specific code already built around raw PyTorch modules or JAX's functional style — Keras's layer and model abstraction sits on top of that and adds a translation layer you may not need. Also skip it for anything beyond inference on OpenVINO, since that backend can't train.

Related repositories