AI Translator
Translate between English and 10 languages, in your browser, with Helsinki-NLP opus-mt models (about 120 MB per language pair, downloaded once).
Why a Translator That Runs in Your Browser
Online translators have an obvious privacy problem: every sentence you translate goes through someone else's server. For confidential documents, contracts, private correspondence, or anything else you would not publish, that round-trip is a real concern.
This tool uses Helsinki-NLP's opus-mt models, converted to ONNX for the browser by Xenova, and runs them with transformers.js on WebAssembly. Each direction is its own model, downloaded only when you pick it: 112 to 137 MB depending on the pair (8-bit quantized files plus the tokenizer), then cached by your browser. Switching from Spanish to French downloads the French model too.
Seventeen of the models are Apache 2.0. Three (English to German, Russian to English and Chinese to English) are CC BY 4.0, which also allows commercial use but asks for credit: they are by Helsinki-NLP (Jörg Tiedemann and the OPUS-MT team, University of Helsinki), converted by Xenova.
Languages: Spanish, French, German, Italian, Portuguese (Brazilian), Russian, Chinese (Simplified), Japanese, Arabic and Dutch, in both directions with English, except English to Japanese. The only commercially licensed browser model for that direction, opus-mt-en-jap, is trained on Bible text and turned an ordinary meeting note into nonsense in our test, so it is not offered. Portuguese uses Helsinki's multi-language ROMANCE models, because no dedicated English-Portuguese opus-mt model exists.
How the Translator Works
Pick a source and a target language, click Load model (or just Translate, which loads it), and paste your text. The tool keeps your paragraph breaks, splits each paragraph into sentences (including Chinese, Japanese and Arabic sentence marks) and translates them one at a time, so long documents work without a length limit. Some models need a target-language tag in front of each sentence, which the tool adds for you: >>pt_BR<< for Portuguese, >>cmn_Hans<< for Chinese and >>ara<< for Arabic.
We ran a two-sentence test passage through all 19 offered directions with transformers.js 3.0.2 (October 2026): every one produced a correct translation, in about 1 to 2 seconds on a desktop CPU after the model had loaded. Results on idioms and specialist vocabulary are weaker than on plain prose.
opus-mt models are bilingual, one model per direction, and the tool translates only to or from English. To go from Spanish to French, translate Spanish to English, then paste the result back in and translate English to French; expect errors to compound.
Frequently Asked Questions
Built by Derek Giordano · Part of Ultimate Design Tools