Translate text — 100% private, in your browser
Privacy-first
Your text never leaves your device
Works offline
After first model download
Free, no signup
No quota, no per-character pricing
Related Tools
Most online translators — Google Translate, DeepL, Bing — ship every sentence you type to a server before showing you a translation. That's fine for restaurant menus and casual messages, but it's a hard no for legal text, medical records, employment contracts, internal company documents, or anything covered by an NDA. FormatFuse takes the opposite approach: the translation model itself is downloaded once into your browser and every word you type stays on your device.
We run Helsinki-NLP's open opus-mt models through transformers.js, accelerated by WebGPU when your browser supports it (Chrome, Edge, modern Safari) and falling back to WebAssembly on everything else. The first time you pick a language pair, your browser fetches that pair's model from Hugging Face's CDN — typically 60–80 MB — and caches it. Every translation after that is offline and instant. There's no signup, no API quota, and no "DeepL Pro is $9/month" upsell. The whole tool is free to use, forever, with no usage cap.
v0 of the translator handles 16 languages and gives you direct translation between any of them and English. Non-English to non-English (Spanish to Japanese, for example) is on the roadmap for v1, where we're integrating a higher-quality model that supports cross-language pairs without pivoting through English. PDF translation — keeping the original layout, fonts, and tables — is also coming as v1, building on this same on-device translation engine.
Quality-wise, opus-mt is honest. It's good enough for emails, chat messages, technical documentation, code comments, and informal text. It's not as polished as DeepL on long literary passages or as fluent as GPT-4 on idiomatic English. Pick FormatFuse when privacy and offline access matter more than the last 5% of fluency — and pick a server-based tool when you're translating marketing copy that needs to sing.
Translate text — 100% private, in your browser — Frequently Asked Questions
Is my text actually private?
Yes. The translation model runs locally inside your browser via transformers.js — the same engine our background remover and OCR use. You can verify this in your browser's Network tab: after the first model download, no outbound requests are made when you translate. Your text never leaves your device, and we don't have a server that could log it even if we wanted to.
Which language pairs are supported in v0?
Any of 16 languages → English, and English → any of those 16 languages. The list: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Japanese, Chinese, Arabic, Korean, Hindi, Swedish, Turkish, Finnish. Non-English to non-English (e.g. French → Japanese) isn't directly supported in v0 — you can pivot through English in two steps for now, and v1 will add direct cross-language pairs.
How does quality compare to DeepL or Google Translate?
Honestly: server-based tools have an edge on long, idiomatic, or literary text. DeepL is particularly strong on European languages and Google is strongest on Asian-language fluency. opus-mt (the Helsinki-NLP models we run) holds its own on technical text, short sentences, structured documents, and most everyday translations. We trade a small amount of fluency for full privacy and unlimited free usage — the right call for sensitive content. v1 will add a higher-quality model option for users who need DeepL-tier output.
Why does it download a large model on first use?
Each language pair has its own opus-mt model — typically 60–80 MB after quantization. That model is fetched from Hugging Face's CDN the first time you select that pair, and your browser caches it. Switch back to a previously-used pair later and it loads instantly from cache. The download is one-time per pair and happens entirely between you and the CDN — FormatFuse never sees the request.
Does it work offline?
Yes, after the first model download. Once the pair you want is cached, you can disconnect from the internet and keep translating. The whole pipeline — tokenizer, model, decoder — runs inside your browser. This makes the tool usable on flights, in secure environments, and anywhere a server-based translator would simply fail.
How is FormatFuse cheaper than DeepL Pro at $9/month?
Because we don't run servers. DeepL Pro pays for GPUs that translate everyone's text on demand; that cost has to be passed on. FormatFuse downloads the model to your machine once, then your machine does the work. There's nothing to subscribe to, no quota, and no per-character pricing.
Will FormatFuse translate PDFs, Word docs, or subtitles?
Not yet — v0 is a text-in / text-out translator. PDF translation (preserving the original layout, fonts, and tables) is the headline feature for v1, and it builds directly on the engine you're using here. Subtitle file translation (.srt, .vtt) is also planned. For now, paste text into the box and we'll translate it.
Is there a character limit?
No hard limit, but translation gets slower as input grows. We show a warning above 5,000 characters because opus-mt processes text in segments and very long inputs may take a few seconds. For book-length text, paste a chapter at a time.
Is my text actually private?
Yes. The translation model runs locally inside your browser via transformers.js — the same engine our background remover and OCR use. You can verify this in your browser's Network tab: after the first model download, no outbound requests are made when you translate. Your text never leaves your device, and we don't have a server that could log it even if we wanted to.
Which language pairs are supported in v0?
Any of 16 languages → English, and English → any of those 16 languages. The list: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Japanese, Chinese, Arabic, Korean, Hindi, Swedish, Turkish, Finnish. Non-English to non-English (e.g. French → Japanese) isn't directly supported in v0 — you can pivot through English in two steps for now, and v1 will add direct cross-language pairs.
How does quality compare to DeepL or Google Translate?
Honestly: server-based tools have an edge on long, idiomatic, or literary text. DeepL is particularly strong on European languages and Google is strongest on Asian-language fluency. opus-mt (the Helsinki-NLP models we run) holds its own on technical text, short sentences, structured documents, and most everyday translations. We trade a small amount of fluency for full privacy and unlimited free usage — the right call for sensitive content. v1 will add a higher-quality model option for users who need DeepL-tier output.
Why does it download a large model on first use?
Each language pair has its own opus-mt model — typically 60–80 MB after quantization. That model is fetched from Hugging Face's CDN the first time you select that pair, and your browser caches it. Switch back to a previously-used pair later and it loads instantly from cache. The download is one-time per pair and happens entirely between you and the CDN — FormatFuse never sees the request.
Does it work offline?
Yes, after the first model download. Once the pair you want is cached, you can disconnect from the internet and keep translating. The whole pipeline — tokenizer, model, decoder — runs inside your browser. This makes the tool usable on flights, in secure environments, and anywhere a server-based translator would simply fail.
How is FormatFuse cheaper than DeepL Pro at $9/month?
Because we don't run servers. DeepL Pro pays for GPUs that translate everyone's text on demand; that cost has to be passed on. FormatFuse downloads the model to your machine once, then your machine does the work. There's nothing to subscribe to, no quota, and no per-character pricing.
Will FormatFuse translate PDFs, Word docs, or subtitles?
Not yet — v0 is a text-in / text-out translator. PDF translation (preserving the original layout, fonts, and tables) is the headline feature for v1, and it builds directly on the engine you're using here. Subtitle file translation (.srt, .vtt) is also planned. For now, paste text into the box and we'll translate it.
Is there a character limit?
No hard limit, but translation gets slower as input grows. We show a warning above 5,000 characters because opus-mt processes text in segments and very long inputs may take a few seconds. For book-length text, paste a chapter at a time.