TranslateGemma

English | 简体中文 | 繁體中文 | 日本語

TranslateGemma

🌍 Local AI Translation Service with Web UI, REST API & MCP Integration
55 languages, smart chunking, streaming output. Powered by Google's TranslateGemma.

Docker Python 3.11+ License: MIT CUDA 12.4

---

✨ Highlights

• 🌐 Web UI - Beautiful, responsive translation interface
• 🔌 REST API - Full-featured API with streaming support
• 🤖 MCP Integration - Works with Claude Desktop & AI assistants
• 🌍 55 Languages - Full TranslateGemma language support
• 📚 Smart Chunking - Handle unlimited text length (chunk_size=100)
• ⚡ Streaming Output - Real-time translation progress
• 🐳 All-in-One Docker - 82GB image with all 6 models embedded
• 🎯 Multi-Model - 4B/12B/27B with Q4/Q8 quantization

---

🎬 Quick Start

Option 1: Docker All-in-One (Recommended)

# Pull the all-in-one image (82GB, includes all models)
docker pull neosun/translategemma:v1.0.0-allinone

# Run with GPU
docker run -d --gpus '"device=0"' \
  -p 8022:8022 \
  -e MODEL_NAME=27b \
  -e QUANTIZATION=8 \
  --name translategemma \
  neosun/translategemma:v1.0.0-allinone

# Access Web UI
open http://localhost:8022

Option 2: Docker with Model Download

# Pull lightweight image (10GB)
docker pull neosun/translategemma:latest

# Run (models download on first use)
docker run -d --gpus '"device=0"' \
  -p 8022:8022 \
  -v ~/.cache/translate/models:/root/.cache/translate/models \
  --name translategemma \
  neosun/translategemma:latest

Option 3: Docker Compose

# docker-compose.yml
services:
  translategemma:
    image: neosun/translategemma:v1.0.0-allinone
    container_name: translategemma
    ports:
      - "8022:8022"
    environment:
      - NVIDIA_VISIBLE_DEVICES=0
      - MODEL_NAME=27b
      - QUANTIZATION=8
      - BACKEND=gguf
      - GPU_IDLE_TIMEOUT=0
      - MAX_CHUNK_LENGTH=100
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["0"]
              capabilities: [gpu]

docker-compose up -d

---

🖥️ Web UI

Access the beautiful web interface at http://localhost:8022:

Features:

• 🎨 Dark/Light theme toggle
• 🔄 Language swap button
• 📊 Real-time translation stats
• ⚙️ Advanced parameter controls
• 📁 File upload support
• 🔥 GPU status monitoring

---

🔌 REST API

Translate Text

# Simple translation
curl -X POST http://localhost:8022/translate \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "target_lang": "zh"}'

# Response
{
  "translation": "你好，世界",
  "source_lang": "en",
  "target_lang": "zh",
  "model": "27b-Q8",
  "time_ms": 1234
}

Streaming Translation

curl -X POST http://localhost:8022/translate/stream \
  -H "Content-Type: application/json" \
  -d '{"text": "Long text here...", "target_lang": "zh"}'

API Endpoints

Endpoint	Method	Description
/translate	POST	Translate text
/translate/stream	POST	Streaming translation
/config	GET	Get current config
/models	GET	List available models
/languages	GET	List supported languages
/gpu/status	GET	GPU memory status
/health	GET	Health check

---

⚙️ Configuration

Environment Variables

Variable	Default	Description
MODEL_NAME	27b	Model size: 4b, 12b, 27b
QUANTIZATION	8	Quantization: 4 or 8
BACKEND	gguf	Backend: gguf, pytorch
GPU_IDLE_TIMEOUT	0	Auto-unload timeout (0=immediate)
MAX_CHUNK_LENGTH	100	Safe chunk size for completeness
DEFAULT_OVERLAP	0	Sliding window overlap (0=disabled)
NVIDIA_VISIBLE_DEVICES	0	GPU device ID

Model Selection Guide

Model	VRAM	Quality	Speed	Use Case
4B-Q4	~3GB	Good	Fast	Quick translations
4B-Q8	~5GB	Better	Fast	Daily use
12B-Q4	~7GB	High	Medium	Balanced
12B-Q8	~12GB	Higher	Medium	Recommended
27B-Q4	~15GB	Best	Slow	High quality
27B-Q8	~28GB	Best+	Slow	Professional ⭐

---

🌍 Supported Languages (55)

Code	Language	Code	Language	Code	Language
en	English	zh	Chinese (Simplified)	zh-TW	Chinese (Traditional)
ja	Japanese	ko	Korean	yue	Cantonese
fr	French	de	German	es	Spanish
pt	Portuguese	ru	Russian	ar	Arabic
hi	Hindi	th	Thai	vi	Vietnamese

...and 40 more languages. See /languages endpoint for full list.

---

🐳 Docker Images

Image	Size	Description
neosun/translategemma:v1.0.0-allinone	82GB	All 6 models embedded
neosun/translategemma:latest-allinone	82GB	Latest all-in-one
neosun/translategemma:v1.0.0	10GB	Lightweight, download models
neosun/translategemma:latest	10GB	Latest lightweight

---

📊 Technical Details

Smart Chunking

TranslateGemma truncates long texts. Our smart chunking ensures complete translations:

chunk_size=100  →  100% translation completeness
chunk_size=120  →  44% completeness (truncation)
chunk_size=150  →  44% completeness (truncation)

Key Finding: chunk_size=100 is the safe boundary for complete translations.

Context Consistency

TranslateGemma automatically maintains consistency across chunks:

• ✅ Pronouns (He/She/They)
• ✅ Terminology (NLP, AI)
• ✅ Proper nouns (Google, Microsoft)
• ✅ Gender consistency

No overlap needed for context preservation.

---

🛠️ Local Development

Prerequisites

• Python 3.11+
• NVIDIA GPU with CUDA 12.4+
• 16GB+ VRAM (for 27B model)

Installation

# Clone repository
git clone https://github.com/neosun100/translategemma.git
cd translategemma

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -e ".[cuda]"

# Run server
uvicorn app_fastapi:app --host 0.0.0.0 --port 8022

---

🤖 MCP Integration

Use with Claude Desktop or other MCP-compatible AI assistants:

{
  "mcpServers": {
    "translategemma": {
      "command": "python",
      "args": ["mcp_server.py"],
      "env": {
        "TRANSLATEGEMMA_URL": "http://localhost:8022"
      }
    }
  }
}

---

📁 Project Structure

translategemma/
├── app_fastapi.py          # FastAPI server
├── mcp_server.py           # MCP server
├── templates/
│   └── index.html          # Web UI
├── static/
│   ├── app.js              # Frontend JS
│   └── style.css           # Styles
├── translategemma_cli/     # Core library
│   ├── translator.py       # Translation logic
│   ├── chunker.py          # Text chunking
│   ├── model.py            # Model loading
│   └── config.py           # Configuration
├── Dockerfile              # Standard image
├── Dockerfile.allinone     # All-in-one image
├── docker-compose.yml      # Compose config
└── tests/                  # Test suite

---

🔬 Research Findings

Our extensive testing revealed:

Parameter	Optimal Value	Reason
chunk_size	100	100% translation completeness
overlap	0	TranslateGemma auto-maintains context
quantization	Q8	Best quality/speed balance

See CHUNKING_RESEARCH_REPORT.md for details.

---

📄 License

MIT License - see LICENSE file.

Note: TranslateGemma models are subject to Google's model license.

---

🙏 Acknowledgments

• Google TranslateGemma - Base translation model
• llama-cpp-python - GGUF inference
• FastAPI - Web framework

---

⭐ Star History

Star History Chart

📱 关注公众号