Transformer-based NMT models, including BERT-like pretraining, mBART, and mT5, offer advantages over RNN/CNN like faster training and better context capture. Practical deployments use techniques like distillation and quantization.