Private text translation, running on your own computer.
Português (Brasil) · Quick start · Language management · Screenshots · API
Local Translator Studio is a native Python application with a clean English web interface, automatic translation, light/dark themes and a local HTTP API. It runs CTranslate2 models directly on the CPU, with SentencePiece or BPE tokenizers. After installation, translation works offline.
Version 2.2.0 · Python 3.10–3.12 x64 · Windows / Linux · AGPL-3.0
- Automatic translation after a 600 ms typing pause, including paste, language changes and IME composition.
- Detect language → English defaults, with configurable source and target languages.
- English as the bridge between two other languages; optional intermediate text display.
- Web language management: search, install, repair, remove, download progress and operation logs.
- Equivalent Windows CMD and Linux CLI commands for model management.
- Persistent light/dark themes and responsive English pages.
- Local processing, no API key and no daily translation quota.
- A compatible text endpoint for Hybrid Translator.
Install 64-bit Python 3.10, 3.11 or 3.12. No Docker, GPU or PowerShell is required.
Open CMD in the application directory and run:
start.batInstall Python and its matching venv package, then run from the application directory:
bash start.shOpen http://localhost:5000. First startup creates an isolated .venv, installs dependencies and downloads missing models. Keep the terminal open; Ctrl+C stops the server. Internet is required for setup and new downloads, but installed translation routes work offline.
In the Tools release ZIP, the application directory is Tools/LocalTranslator. In a standalone GitHub checkout, it is the repository root.
If port 5000 is busy, use start.bat --port 5001 or bash start.sh --port 5001, then open http://localhost:5001.
Open Models:
- Search for a language by name or code.
- Select Install to download and validate both directions, or Repair for a configured language with missing weights.
- Follow the progress indicator, transferred bytes and operation log. Keep the server running until completion.
- Select Remove, then confirm in the dialog to delete a language's two models and cached archives. Cancel leaves the collection untouched.
New languages become available to the server immediately after the job completes. Reopen or reload an existing Translate tab to refresh its language selectors. The Models page updates automatically without a server restart. Removed default languages remain removed on subsequent startup. English is required because it connects the other languages.
One model operation runs at a time. Removal waits for active translations before unloading weights. A failed optional pair remains disabled; other successful installations are preserved. Use Refresh catalogue to retrieve the latest official model list.
Stop the server before managing models from another terminal. Run inside the application directory:
start.bat --list-languages
start.bat --install-languages de it ko
start.bat --remove-languages de it
start.bat --install-all-languages
start.bat --list-languages --refresh-catalog
start.bat --setup-only
start.batbash start.sh --list-languages
bash start.sh --install-languages de it ko
bash start.sh --remove-languages de it
bash start.sh --install-all-languages
bash start.sh --list-languages --refresh-catalog
bash start.sh --setup-only
bash start.shde it ko adds German, Italian and Korean. Substitute any listed codes; use spaces to select several. Management commands exit when finished. --setup-only repairs the configured collection without starting the server; it does not reinstall intentionally removed languages. --skip-install starts an already configured runtime. The same options are available through python run.py.
Installing the entire catalogue can take several gigabytes and significant download time. Archive compatibility is validated during installation; not every optional route has been tested in this project. Repeat an interrupted operation to reuse completed downloads. Installation preserves other languages; removal also deletes the selected language's cached archives.
The default installation provides 18 models / 10 languages including English:
| Language | API code | Language | API code |
|---|---|---|---|
| English | en |
Russian | ru |
| Chinese (Simplified) | zh-Hans |
Portuguese | pt |
| Chinese (Traditional) | zh-Hant |
Spanish | es |
| Vietnamese | vi |
French | fr |
| Thai | th |
Japanese | ja |
The bundled official catalogue snapshot, dated 2026-10-04, lists 49 non-English bidirectional pairs. Only pairs with models both to and from English are offered. Choices include German, Italian, Korean, Arabic, Hindi, Ukrainian and a separate Brazilian Portuguese model (pb). Check the list command or Models page for the current catalogue.
Aliases zh, zt, zh-CN, zh-TW, pt-BR and ptbr remain compatible with the original models. pt-BR maps to generic pt; select optional pb explicitly for the Brazilian model. Automatic detection can be ambiguous for short text or scripts shared by languages; choose the source explicitly when needed.
Model quality varies. Passing through English may compound errors, and a successful model-load test does not establish translation accuracy.
Real captures of the running application. All pages are English; both themes are included.
| Page | Purpose | Light | Dark |
|---|---|---|---|
| Translate | Automatic text translation | View | View |
| Models | Install, repair and remove languages | View | View |
| API Guide | Endpoint documentation and request tester | View | View |
| About | Engine, privacy and integration | View | View |
| Licenses | Third-party credits | View | View |
| Error page | English error messages | View | View |
The server binds to 127.0.0.1. No API key is needed for translation.
| Endpoint | Method | Purpose |
|---|---|---|
/translate |
POST | Translate one string or a JSON list |
/detect |
POST | Detect a configured source language |
/languages |
GET | Configured languages and targets |
/health |
GET | Server availability and model readiness |
/studio/status |
GET | Installed/missing models and limits |
/frontend/settings |
GET | Language defaults and request limits |
import json
from urllib.request import Request, urlopen
request = Request(
"http://localhost:5000/translate",
data=json.dumps({"q": "Bonjour tout le monde", "source": "auto", "target": "en"}).encode(),
headers={"Content-Type": "application/json"},
)
with urlopen(request) as response:
print(json.load(response)["translatedText"])Requests support plain text, up to 30 items and 64,000 characters total, with a maximum 1 MiB body. JSON lists return lists of translations; form fields also work. HTML/file translation is not implemented. Open /docs for examples and a local request tester.
For Hybrid Translator, configure a provider accepting a local /translate endpoint at http://localhost:5000. Enable fallback manually in the extension when you want to use it. This app does not modify extension dictionaries or settings.
The model-management endpoints are intended for the local Models page. Writes require its current page token and same-origin requests; they are not enabled through translation API CORS.
| Layer | Implementation |
|---|---|
| Runtime and startup | Python, isolated venv, BAT / shell launchers |
| HTTP service | Flask and Waitress, localhost only |
| Neural translation | CTranslate2 on CPU; up to three cached models |
| Tokenization | SentencePiece, or BPE with subword-nmt and sacremoses |
| Language detection | langdetect with script hints |
| Interface | Jinja templates, local CSS and vanilla JavaScript |
| Model operations | Background worker, cross-process lock, atomic configuration writes |
Exact dependencies are pinned in requirements.txt. Bundled fonts and assets keep the interface independent of external CDNs. Translation text is processed locally and is not saved by the application. Explicit model-install/refresh operations use the internet.
Default downloads use bundled SHA-256 receipts. Optional archives without bundled hashes record the first official HTTPS download's hash for subsequent integrity checks; that local receipt is not independent publisher verification. Original weights/tokenizer notices are preserved. Catalogue source: Argos model index.
Allow several GB of disk space for default archives, extracted weights and dependencies; more for additional languages. 8 GB RAM is a practical starting point, not a universal minimum. No GPU is required.
Application source and interface: AGPL-3.0. Model weights, dependencies and fonts retain their own licenses; see THIRD-PARTY-NOTICES.md, model-notices and web/fonts. The footer offers a source ZIP excluding local models and environments.
Related project: Hybrid Translator.