Firefox's translation feature was built on Mozilla's Bergamot collab with Europe. It's an NMT as you note but it's also a set of language models trained on large language corpora using, among other techniques, the transformer architecture, so people calling it LLM translation don't bother me at all, and I'm the guy that brought it to Firefox (with my architect partner Andre Natal). The best article is probably https://hacks.mozilla.org/2022/06/training-efficient-neural-... though that was before we started integrating it into Firefox proper.

Thanks for the work you do, and your reply, very interesting!