Keras was created by Francois Chollet and first published to GitHub on March 27, 2015. It is a high-level deep learning API whose tagline, “Deep Learning for humans,” captures its purpose: to let people build and train neural networks with a small amount of clear, readable code rather than wrestling with low-level tensor operations.
Keras did not implement the heavy numerical computation itself; it ran on top of a backend engine. At first that backend was Theano, then TensorFlow, and the design deliberately abstracted the underlying framework away. Its concise model-building syntax, where layers are stacked in a few lines, made it the on-ramp through which many practitioners learned deep learning. In 2017 Google made Keras the official high-level API of TensorFlow. With Keras 3, released in 2023, the framework became multi-backend again, able to run on JAX, TensorFlow, or PyTorch and let users pick whichever is fastest for a given model.
Keras matters because it lowered the barrier to deep learning. Much of the early democratization of neural networks, in courses, tutorials, and first prototypes inside companies, happened through its friendly interface.