🌍 Tencent's offline AI translator — 33 languages on your phone, 440 MB, no internet

[Hannah]

A 33-language translator that lives in your phone — and hates the internet

Hy-MT2 by Tencent: the whole translator is one 440 MB file — drop it into a free app, turn on airplane mode, translate forever · open source, Apache 2.0, 1.7K

your Wi-Fi dies. your SIM dies. your translator keeps working — 33 languages, zero bars

you, offline, in a Tokyo basement at 2 AM:
すし を 食べたい です

Hy-MT2, same second, still zero bars:
I want to eat sushi

that’s the whole trick — the model file (440 MB, smaller than a mobile game) lives on your phone, sees your screen, and answers in any of 33 languages

the stamps it obeys — say it once, it remembers the rule

• glossary stamp — “always translate 味噌 as miso, never as soybean paste” — your terms, locked
• style stamp — “translate like a pirate” / “keep it formal” / “12-year-old can read this”
• code stamp — variables, placeholders and tags pass through untouched, only human words get translated
• format stamp — line breaks, separators and layouts come back exactly as they went in

boarding pass — how you fly it

	your flight plan
luggage	grab the Hy-MT2-1.8B GGUF — 440 MB from Hugging Face · repo: github.com/tencent-hunyuan/Hy-MT2
gate	drop it into PocketPal AI (Android + iPhone) · Maid or LM Playground (Android)
mode	airplane — one test online, then never needs a bar again
destination	33 languages · WMT26 subtitle-translation official partner

▸ one minimized window (click to open)

• the myth-busts — DeepL has no offline mode (every word goes to the cloud) · NLLB-200 covers 200 languages but forbids commercial use · Google offline packs are great for camera/menus — install both, use each for what it’s best at
• license, checked — Apache 2.0 straight from the repo’s LICENSE.txt: commercial use, modification, redistribution — all allowed
• the subtitle flight — Tencent officially partners with WMT26’s video-subtitle task; keep timestamps outside the model, translate the text inside
• the honest stamp — “beats Microsoft/Doubao” is a vendor benchmark, not an independent audit · ~3–5 sec per step · quantization can soften tricky text

─ the passport control, one line

the superpower — private text never leaves your phone (journals, drafts, game strings) · 440 MB is the file, not the RAM it needs · never trust any AI translation for legal/medical/safety-critical

The best translator is the one that works when everything else doesn’t.