Fairseq is a sequence-to-sequence learning toolkit developed by Facebook AI Research for Neural Machine Translation. It implements convolutional NMT models as described in two academic papers along with standard LSTM-based models. The toolkit supports multi-GPU training on a single machine and provides fast beam search generation on both CPU and GPU. Pre-trained models are available for English-French, English-German, and English-Romanian translation. New development has moved to the PyTorch version (fairseq-py), and this Lua version is preserved without active support.