Babelo

Offline neural machine translation with local-only, on-device inference. Supports multiple translation models:

• NLLB-200: Meta's No Language Left Behind model (200+ languages, ~2.5GB)
• TranslateGemma: Google's lightweight translation model based on Gemma 3 (55 languages, ~8GB) - Requires a HuggingFace token

Features

• Multiple Models: Choose between NLLB-200 (200+ languages) or TranslateGemma (55 languages)
• Privacy-First: Local translation inference, no data sent to external servers
• Text-to-Speech: Listen to translations in supported languages in the browser
• Saved Translations: Save and revisit your favorite translations
• CLI & Web Interface: Use via terminal or web app
• Fast & Efficient: Optimized models for speed

Tech Stack

• Backend: Python, FastAPI, HuggingFace Transformers
• Frontend: React, TypeScript, Tailwind CSS
• Models: Facebook NLLB-200-distilled-600M, Google TranslateGemma-4b-it

Prerequisites

• Python 3.11+
• uv - Python package installer and resolver
• ~3GB disk space for NLLB-200, ~8GB for TranslateGemma
• For TranslateGemma: Hugging Face account and token (see below)

Setup

1. Install dependencies:

   uv sync

2. Download a translation model:

   NLLB-200 (default, ~2.5GB):

   uv run bab download

   TranslateGemma (~8GB):

   uv run bab download --model translategemma

   Note: TranslateGemma requires accepting the license on Hugging Face and setting the HF_TOKEN environment variable. Visit https://huggingface.co/google/translategemma-4b-it to accept the license.

3. Verify the model files:

   uv run bab verify
   # Or for a specific model:
   uv run bab verify --model translategemma

4. Run the development server:

   uv run fastapi dev server/main.py

5. Open http://localhost:8000 in your browser to see the API.

Available Models

Model	Languages	Size	Auth Required
nllb	200+	~2.5GB	No
translategemma	55	~8GB	Yes (HF token)

List all models with their download status:

uv run bab models

CLI Commands

The bab CLI provides commands for managing translation models:

List available models

uv run bab models

Check model status

uv run bab status
uv run bab status --model translategemma

Shows whether the model is downloaded and loaded, along with storage location and size.

Download a model

uv run bab download
uv run bab download --model translategemma

Downloads the model to the local cache. Use --force to re-download:

uv run bab download --force

Verify model files

uv run bab verify
uv run bab verify --model translategemma

Checks that all required model files are present.

List supported languages

uv run bab languages
uv run bab languages --model translategemma

Shows all supported languages for the specified model.

Test model loading

uv run bab load-test
uv run bab load-test --model translategemma

Loads the model into memory to verify it works correctly.

Translate text

uv run bab translate "Hello, world!" -s eng_Latn -t fra_Latn
uv run bab translate "Hello, world!" -s en -t fr --model translategemma

Note: Language codes differ between models. Use bab languages --model <id> to see available codes.

Delete a model

uv run bab delete
uv run bab delete --model translategemma

Custom cache directory

All commands support a custom cache directory:

uv run bab --cache-dir /path/to/cache status

Interactive mode

Run without arguments to enter interactive mode:

uv run bab

Commands in interactive mode:

• /model - Select translation model
• /source - Set source language
• /target - Set target language
• /swap - Swap languages
• /status - Show status
• /languages - List languages
• /models - List models
• /help - Show help

API Documentation

Once the server is running, visit:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

Key API Endpoints

Models:

• GET /model/list - List available models
• GET /model/list/status - Get status of all models
• GET /model/status?model_id=nllb - Get status of a specific model
• POST /model/download?model_id=nllb - Download a model
• POST /model/remove?model_id=nllb - Remove a model

Translation:

• POST /translate - Translate text

  {
    "text": "Hello",
    "source_language_code": "eng_Latn",
    "target_language_code": "fra_Latn",
    "model_id": "nllb"
  }

Languages:

• GET /languages?model_id=nllb - Get supported languages

TranslateGemma Setup

TranslateGemma requires authentication with Hugging Face:

1. Create a Hugging Face account at https://huggingface.co

2. Visit https://huggingface.co/google/translategemma-4b-it and accept the license

3. Create an access token at https://huggingface.co/settings/tokens

4. Set the HF_TOKEN environment variable:

   export HF_TOKEN=hf_your_token_here
   uv run bab download --model translategemma

Project Structure

babelo/
├── cli/
│   ├── cli.py           # CLI commands
│   └── interactive/     # Interactive mode
├── core/
│   ├── model.py         # Model registry and backends
│   ├── languages.py     # Language code mappings
│   ├── database.py      # Saved translations storage
│   └── preferences.py   # User preferences
├── frontend/
│   └── src/
│       ├── components/
│       │   ├── ModelSelector.tsx
│       │   └── ...
│       ├── hooks/
│       └── utils/
├── server/
│   └── routes/
│       ├── model.py     # Model management endpoints
│       ├── translate.py # Translation endpoint
│       ├── languages.py # Languages endpoint
│       └── ...
├── tests/
├── README.md
└── pyproject.toml

Development

Running tests

uv run pytest

With verbose output:

uv run pytest -v

Adding dependencies

uv add <package-name>

Adding dev dependencies

uv add --dev <package-name>

Model Information

NLLB-200

NLLB-200-Distilled-600M from Meta AI.

• Model size: ~2.5GB
• Supported languages: 200+ languages
• Architecture: Encoder-Decoder Transformer
• Language codes: BCP-47 style (e.g., eng_Latn, fra_Latn)

TranslateGemma

TranslateGemma-4b-it from Google.

• Model size: ~8GB
• Supported languages: 55 languages
• Architecture: Gemma 3 based multimodal model
• Language codes: ISO 639-1 (e.g., en, fr, de-DE)
• License: Requires accepting Gemma license on Hugging Face