One of four (MXNet and CNTK alongside TF and Theano), and the Amazon deep learning API forked Keras to default to MXNet support before it was really ready - which irked the Keras authors quite a bit.