Keras

Keras is an open-source Python deep learning framework for developers and researchers who want to build and train models without tying their code to one backend. It runs locally on Linux and macOS, with WSL2 recommended for Windows. Its Apache 2.0 license permits commercial use and modification.

The same modeling API works with JAX, TensorFlow and PyTorch. OpenVINO supports inference only. This gives teams room to choose a backend for its performance or deployment ecosystem while keeping compatible model code. Keras supports CPU execution and NVIDIA GPUs through CUDA, and training can scale from a laptop to GPU or TPU clusters.

Built-in training and evaluation methods reduce the amount of routine code developers need to write. The Functional API supports connected model architectures, while custom layers and model subclasses allow more control. Models can also participate in native backend training loops and use existing TensorFlow dataset pipelines or PyTorch DataLoaders, regardless of the chosen training backend.

Keras covers computer vision, language and audio processing, alongside time-series forecasting and recommender systems. KerasHub supplies model architectures and pretrained checkpoints available through Kaggle Models for training and inference. These include Gemma, Qwen, GPT-OSS and Mistral language models, plus Stable Diffusion for text-to-image generation.