Translate between languages without sending text anywhere.
Real translation, zero upload.
Every character stays on your device. Models run through Transformers.js with WebGPU acceleration and a WASM fallback.
100% private
Your text never leaves the browser. Translation runs locally, so nothing is uploaded to a server or stored in the cloud.
WebGPU acceleration
Inference uses your GPU when available and silently falls back to WASM on browsers without WebGPU — same result, different speed.
30+ languages
English, Chinese, Japanese, Korean, Spanish, French, German, and more. Pairs without a direct model pivot through English automatically.
Auto-detect
Leave the source on Auto and the tool detects the language for you, then routes to the right on-device model.
Bidirectional swap
Flip source and target with one tap. Your text and results swap instantly so translating the reply is effortless.
Local history
Recent translations are kept on your device. Reload any of them to edit, copy, or download — nothing is synced to a server.
Paste, pick, translate.
- 01
Paste your text
Drop in a sentence, paragraph, or whole passage. Set the source language, or let it auto-detect.
- 02
Choose a target
Pick from 30+ languages. If no direct model exists, translation routes through English automatically.
- 03
Read or export
Get the result on screen instantly, copy it, or download a .txt file — all on your device.
About Translate
Translate is a private, on-device translator built on Meta AI's M2M-100 model (MIT-licensed, many-to-many) running through Transformers.js. Paste text, pick a language pair, and the result appears in your browser — no server, no upload, no account. It supports 30+ languages and translates directly between any pair — even two non-English languages — with no English pivot.
Everything runs locally. Each language model downloads once from Hugging Face and caches in your browser; after that, translation works offline. Because the bytes never leave your device, the tool is safe for confidential documents and fine for commercial use.
How it works
Pick a source language or leave it on Auto-detect. Choose a target from the 30+ supported languages. When no direct model covers a pair — say Estonian to Chinese — the engine routes through English automatically (Estonian→English→Chinese), so you always get a result.
Press Translate and the text is split into manageable chunks, run directly through the single multilingual M2M-100 model on your GPU (WebGPU) or CPU (WebAssembly), then reassembled. You can swap source and target with one tap to translate a reply.
Languages & models
The underlying model is Meta AI's M2M-100: a single multilingual many-to-many network licensed under MIT (commercial use allowed). It runs entirely in the browser, which is why the first run downloads one model (~430 MB, quantized) that is then cached.
Every listed language translates directly to every other — including non-English pairs like Chinese → Japanese — because a single model covers them all.
Privacy & limits
Your text never leaves your device. Open the browser's Network panel and translate with the tab offline — it still works. The only thing cached is the model weights, and they are stored locally, not on our servers.
The free tier translates up to 5,000 characters per run with unlimited runs; Pro raises that to 50,000 characters per run and keeps a longer local history. Both tiers are fully on-device and commercially usable.