Firefox is the only web browser that provides local in-page translation that is private by design. What does it take to train a high-quality Neural Machine Translation model? How do we compress the models from 800 to 20 MB and make them efficient to perform well on a CPU? What infrastructure and scale does it require to run training for many languages? We will answer all these questions and talk about the unique challenges and the future of the project. And yes, the good news: code, training data and models are fully open-source!