A free, private, open-source desktop translator. DeepL-grade quality — but you bring the models, you bring the keys, and your text never leaves your machine unless you decide it should.
English · 简体中文
LLMT (LLM-Translate) is a tiny, fast desktop app for Chinese ⇄ English translation, built as an open-source answer to DeepL. It runs as a native window (Tauri + Rust), installs in ~11 MB, and gives you two ways to translate:
- Dictionary — instant, 100% offline, zero cost.
- LLM — full-sentence and full-document translation through any model you like: a cloud API or a model running on your own laptop.
Where DeepL locks "unlimited" usage and file translation behind a $57.49 / user / month business plan, LLMT lets you do the same work for $0 (offline) or literally pennies (cloud API, pay-as-you-go). More on that below.
| Mode | Best for | Cost | Network |
|---|---|---|---|
| Dictionary | single words & short phrases | free | offline |
| LLM | sentences, paragraphs, whole documents | free (local) / pennies (API) | local or cloud |
-
Dictionary mode is powered by two bundled offline dictionaries:
- CC-CEDICT for Chinese → English (≈120k entries)
- ECDICT for English → Chinese (≈58k common words, with phonetics)
Look up a word like
recastand you instantly get[.ri:'kɑ:st] vt. 重铸, 改造— no API call, no waiting, no bill. -
LLM mode talks to any OpenAI-compatible endpoint, Anthropic, or a local server. Bring your own key, or point it at a model on
127.0.0.1.
Drop a .txt straight onto the window and LLMT will:
- Read the file,
- Translate the whole document with your chosen LLM,
- Pop a native Save dialog so you can export the result as
*_translated.txt.
Long files are automatically chunked along paragraph / sentence boundaries (so you never hit token limits), translated chunk-by-chunk with a live Translating 3/8… progress indicator, and stitched back together.
Click the speaker to hear any text read aloud by Kokoro-82M, a high-quality neural TTS running entirely in the app (WebAssembly) — no server, no Docker. 50+ voices across English and Mandarin.
A warm "ink & seal" interface with a refined type system, light/paper and dark themes, and smooth micro-interactions. Built to feel like a premium tool, not a form with two boxes.
DeepL's comparable tier is Business:
DeepL Business — $57.49 / user / month (billed annually) Unlimited characters · 100 file translations / month · 20 shared style rule lists · unlimited glossaries · translation memory · unlimited saved translations · SCIM · domain capture · 99.0% availability · Write Pro
Here's what the same work costs in LLMT.
| LLMT option | Cost | Notes |
|---|---|---|
| Dictionary | $0 | fully offline, instant |
| Local LLM (e.g. Hy-MT2) | $0 | runs on your machine after a one-time download |
If you only use Dictionary mode or a local model, you never pay anything — and nothing ever leaves your computer.
Assumption: a typical article ≈ 1,000 words ≈ ~1.5K input tokens + ~2K output tokens (translation expands a little into Chinese). Prices are per 1M tokens as of 2026-05-27.
| Model | Input / Output | Per article | 100 articles |
|---|---|---|---|
| DeepSeek V4 Pro | $0.435 / $0.87 | ≈ $0.0024 | ≈ $0.24 |
| MiMo v2.5 Pro | $0.435 / $0.87 | ≈ $0.0024 | ≈ $0.24 |
Both providers now sit at the same low price — and it's permanent, not a limited-time promo.
| DeepL Business | LLMT (cloud API) | |
|---|---|---|
| 100 file translations | $57.49 / mo | ≈ $0.24 |
| Offline / local | ❌ | $0 |
That's roughly 240× cheaper — about a quarter, versus the $57.49 DeepL charges.
No seats. No annual lock-in. No per-character meter. You pay the model provider directly, for exactly what you use — usually fractions of a cent per page.
Want zero cost and zero data leaving your machine? Run a model locally and point LLMT's Local provider at it (any OpenAI-compatible server — llama.cpp, LM Studio, Ollama, vLLM, …).
We especially recommend Tencent's Hy-MT2 family — purpose-built translation models (Apache-2.0, 33 languages):
| Model | Size | Runs on | Notes |
|---|---|---|---|
| Hy-MT2-1.8B | 1.8B | laptops / on-device | ~440 MB quantized, fast |
| Hy-MT2-7B | 7B | a decent GPU / Apple Silicon | great quality/speed balance |
| Hy-MT2-30B-A3B | 30B (MoE, ~3B active) | workstation GPU | top quality |
Tencent reports the Hy-MT2 series outperforms open-source models like DeepSeek-V4-Pro and Kimi K2.6, as well as commercial APIs from Microsoft and Doubao on translation. The 1.8B is small enough to live happily on a regular laptop.
Don't want to manage a model? Just use Dictionary mode — it's already built in and needs nothing at all.
Modern LLMs are excellent translators. They preserve tone, handle idioms, keep formatting, and read naturally — frequently matching or beating dedicated translation services on real-world prose, technical docs, and nuanced passages. With LLMT you get that quality on your terms: pick the exact model you trust, swap it any time, and keep full control of your data and your spend.
Grab the latest installer from Releases:
- Windows —
LLMT_x.y.z_x64-setup.exe(NSIS) orLLMT_x.y.z_x64_en-US.msi - macOS —
LLMT_x.y.z_*.dmg
macOS: the app is unsigned, so on first launch right-click → Open, or run
xattr -dr com.apple.quarantine /Applications/LLMT.app.
- Dictionary works immediately — no setup.
- For LLM mode, open ⚙ Settings and add an API key (OpenAI / Anthropic) or a Local endpoint URL + model name.
- The first time you use voice, the Kokoro model (~100 MB) downloads once, then caches for offline use.
# prerequisites: Node 18+, Rust (stable)
npm install
# regenerate the bundled dictionaries (optional — JSON is committed)
npm run build-dict # CC-CEDICT -> src/dict.json
npm run build-endict # ECDICT -> src/endict.json (needs ecdict.csv, see script header)
# run in dev
npm run tauri dev
# package installers / .dmg
npm run tauri buildThe icon set is generated from src-tauri/icons/app-icon-src.png via node scripts/build-icon.mjs + npx tauri icon.
- Dictionary mode and local LLM mode are fully offline — your text never leaves the device.
- Cloud API mode sends text only to the provider you configured, using your key. LLMT has no servers and collects nothing.
- API keys are stored locally in your OS user-data directory.
- Tauri 2 + Rust backend (native window, ~11 MB installer)
- TypeScript + Vite frontend
- kokoro-js — neural TTS in WebAssembly
- CC-CEDICT + ECDICT — offline dictionaries
| Asset | Source | License |
|---|---|---|
| Chinese→English dictionary | CC-CEDICT | CC BY-SA 4.0 |
| English→Chinese dictionary | ECDICT | free / MIT |
| Neural TTS | Kokoro-82M | Apache-2.0 |
| Code | this project | MIT |
If you redistribute, please keep the CC-CEDICT attribution and share-alike notice.
LLMT — translate everything, own everything.
Made with Tauri, Rust & a lot of care.