AI Translator

Neural machine translation without the cloud

This translator uses Opus-MT, a family of compact neural translation models trained by the University of Helsinki on the open OPUS parallel corpus. Each model is a Marian transformer specialised in one language pair, which keeps it small enough to run inside a web page. The model is downloaded from Hugging Face the first time you use a pair, stored in your browser cache and executed with WebAssembly in a background Web Worker, so the page stays responsive while it works.

Because everything happens locally, you can translate confidential emails, contracts, medical notes or unpublished writing without sending them to a translation service. There is no account, no usage limit and no log of what you typed. Once a model is cached, translating the same pair again starts in seconds.

How long texts are translated

Translation models work best on short passages, so the tool splits your text into sentences and groups them into segments of a few hundred characters. Each segment is translated in turn, and the results are joined back together with your original line breaks and paragraphs preserved. The progress bar shows how many segments are done. When there is no direct model for a pair — for example Polish to Swedish — the tool translates into English first and then into the target language, and tells you which models it will load.

Getting the best translation

Write complete sentences with correct punctuation, since sentence boundaries guide the segmentation. Avoid slang and very long sentences where you can. For names and terms that must not change, check the output carefully. Machine translation is a great first draft, but a human review is still recommended for anything you publish or sign.