A universal language translator that runs entirely inside your browser — no server, no API keys, no data leaving your device.

→ Try it live

Translates text between 55 languages using Meta's NLLB-200 model, loaded directly into the browser via WebAssembly. The model downloads once (~600 MB), is cached locally, and works offline from that point on.

Supported language families

Serve the file from any static server:

python3 -m http.server 8080
# open http://localhost:8080/universal-translator.html

No build step. No dependencies. No backend.

This is the more interesting question.

Privacy. When you paste text into Google Translate, DeepL, or any cloud service, that text travels to a third-party server and may be logged, stored, or used for model training. For casual use, that's usually fine. But what about a contract clause? A medical note? A confidential business document? A personal letter in another language? The moment you translate it via a cloud service, you've handed it to someone else. This tool keeps everything on your device — nothing is transmitted anywhere.

Offline. Download the model once. After that, it works on a plane, in a remote area, or anywhere without a reliable connection. Cloud translators stop working entirely when connectivity drops.

No API costs or rate limits. Production translation APIs charge per character — at scale, this adds up quickly. Local inference has no marginal cost and no usage cap.

Sensitive environments. Air-gapped systems, corporate intranets, and regulated industries (legal, medical, finance) often cannot send data to external APIs at all. A self-contained HTML file that runs locally is deployable anywhere, including places that have never seen the public internet.

Auditability. You can read every line of code in this project. There is no black box, no telemetry, no network calls beyond the one-time model download from Hugging Face.

The trade-off is honest: a ~600 MB one-time download, and quality that is good but not quite at the level of the top commercial services. For most real-world use cases — especially ones where privacy matters — it's a worthwhile exchange.

The download is genuinely one-time. The model is cached in your browser's storage — close the tab, shut down your laptop, come back a week later and it's still there, ready in a few seconds. No re-download. The only way it disappears is if you explicitly clear your browser's site data.

• Transformers.js — runs Hugging Face models in the browser over WebAssembly
• NLLB-200-distilled-600M — Meta's open multilingual model, 200-language coverage, q8-quantised for browser use
• Zero build tooling. One HTML file.

Coloured with china-02 · 米紙 MIZHI — hand-scroll mulberry paper, calligraphy. Warm cream body, indigo translation accent, brick-red brand mark.

Palette pulled from Rangrez, the global colour-palette library that backs all NakliTechie projects.

Built by Chirag Patnaik

Built in a few minutes with Claude.