Keras

Keras 3 separates the modeling API from the math engine underneath: you define layers, models, losses and metrics once, then pick JAX, TensorFlow or PyTorch at startup through the KERAS_BACKEND variable or ~/.keras/keras.json, with OpenVINO as a fourth, prediction-only backend. Functional and subclassing styles both work, fit, evaluate and predict cover the usual loop, and custom train steps or native loops in the chosen framework remain possible. A Keras model can sit inside a PyTorch Module or a JAX function, and it accepts tf.data pipelines or PyTorch DataLoaders on any backend. Most tf.keras code moves over once models are saved in the .keras format. KerasHub adds pretrained architectures and checkpoints, from Gemma language models to Stable Diffusion 3. Install with pip install keras plus at least one backend; Linux and macOS are supported directly, Windows users are pointed to WSL2, and GPU use depends on the backend's own CUDA build. Apache-2.0 and free for commercial use. Suited to practitioners who want readable model code and the freedom to switch frameworks later.