Neural Machine Translation (NMT) is the practice of building and training deep learning models (typically Transformers) that translate text from one language to another. The core architecture: an encoder reads the source language (English) word-by-word, creates a context vector, then a decoder generates the target language (German) word-by-word, using attention to focus on relevant source words at each step. Modern NMT models (mBART, mT5) are pre-trained on 100+ languages, then fine-tuned for specific pairs. Quality depends on: training data size, model capacity, tokenization strategy, and inference-time decoding (beam search vs. greedy).