neural-fortran is a parallel framework for deep learning written in Fortran. It provides implementations of dense, convolutional (1D and 2D), and transformer neural network architectures. The framework includes support for multiple optimizers including SGD, momentum variants, RMSProp, Adagrad, Adam, and AdamW. It offers various activation functions, loss functions, and metrics, and supports loading pre-trained models from Keras HDF5 files. Data-based parallelism is supported for distributed training.