Configure Local Translation

MAITO’s local translation is its standout feature - translate completely offline with complete privacy using AI models running on your own computer.

What You’ll Need

Before configuring local translation:

• Windows x64 system
• Minimum 4GB RAM (8GB+ recommended)
• 2-8GB free disk space for models
• 5-10 minutes for initial model download

Two Ways to Configure

MAITO offers two configuration paths:

Auto Configuration (Recommended)

Perfect for most users:

• MAITO assesses your hardware automatically
• Downloads the best model for your system
• Validates installation
• Runs performance test
• Takes 5-10 minutes total

Manual Configuration

For advanced users who want control:

• Review hardware assessment
• Choose specific model to download
• Configure manually
• Requires understanding of model types

Configuration via Onboarding

If This Is Your First Time

When you first launch MAITO, the onboarding wizard will guide you:

1. Choose Translation Engine → Select “Local AI Translation”
2. Choose Configuration Type → Select “Auto (Recommended)”
3. Device Assessment → MAITO analyzes your hardware
4. Model Download → Automatic download begins
5. Performance Test → 30-second benchmark
6. Complete → Start translating!

Configuration via Settings

If You’re Adding Local Translation Later

If you initially chose DeepL and want to add local translation:

1. Click the gear icon (⚙️) to open Settings
2. Navigate to the Translation tab
3. Select Local On-Device Translation radio button
4. Click Apply Selected Engine
5. The configuration wizard will appear

Understanding Device Assessment

MAITO analyzes your hardware to recommend the optimal model:

Hardware Analyzed

CPU Detection

• Processor model and generation
• Number of cores and threads
• Support for special instructions (AVX2, etc.)

RAM Detection

• Total installed RAM
• Available free RAM
• Memory speed

GPU Detection (if available)

• GPU vendor (NVIDIA, AMD, Intel)
• VRAM amount
• DirectML compatibility

Performance Ratings

MAITO assigns an overall performance rating:

Model Recommendations

Based on your hardware assessment, MAITO recommends a model:

Rosetta Models (Recommended)

MAITO automatically recommends Rosetta models - custom-trained translation models optimized by us:

Rosetta 4B CPU (2.5GB)

• Optimized for CPU-only systems
• Runs on any Windows x64 system
• Best all-around compatibility
• Recommended for most users

Rosetta 4B DML (2.5GB)

• Optimized for DirectML (GPU acceleration)
• Requires compatible GPU
• Faster translation speeds
• Recommended if you have a GPU

Key Features:

• Trained on 32 languages
• Optimized for translation quality
• Small model size (2.5GB)
• Fast inference speeds

Auto Configuration Walkthrough

Let’s walk through the auto-configuration process step-by-step:

Phase 1: Device Assessment (10-30% progress)

MAITO analyzes your hardware:

• Detects CPU capabilities
• Checks available RAM
• Scans for compatible GPUs
• Calculates performance rating

Time: 5-10 seconds

Phase 2: Model Download (30-70% progress)

MAITO downloads the recommended model:

• Selects best model for your hardware
• Downloads from secure servers
• Shows real-time download progress
• Typical size: 2.5GB (Rosetta models)

Time: 3-8 minutes (depending on internet speed)

Phase 3: Model Validation (75-80% progress)

MAITO validates the downloaded model:

• Verifies file integrity
• Tests model can be loaded
• Saves configuration to settings

Time: 10-20 seconds

Phase 4: Performance Test (85-100% progress)

MAITO runs a 30-second benchmark:

• Translates sample text
• Measures translation speed (tokens/second)
• Provides performance rating
• Validates everything works

Time: 30 seconds

After the benchmark, you’ll see the result of the benchmark, to give an indication of what you can expect performance-wise. Not that this does not impact the quality of the translations, only the speed in which translation happens.

What Do Tokens/Second Mean?

You may come across a metric called Tokens/Second when benchmarking AI models. This translates to the following:

• 20+ tokens/sec: Excellent - near-instant translations
• 10-20 tokens/sec: Good - smooth, responsive experience
• 5-10 tokens/sec: Fair - some wait time, but usable
• 2-5 tokens/sec: Poor - noticeable delays
• <2 tokens/sec: Slow - use for short texts only

Manual Configuration

If you choose manual configuration:

Step 1: Review Hardware Assessment

Review the device assessment screen showing your hardware capabilities.

Step 2: Model Selection Screen

Browse available models:

• Model Name: e.g., “Pangaia Software Rosetta 4B CPU”
• Size: e.g., “2.5 GB”
• Compatibility: CPU-only, GPU-accelerated, etc.
• Recommendation: MAITO suggests models with a star icon ⭐

Step 3: Download Model

Select a model and click Download:

• Monitor download progress
• Cancel and retry if needed
• Model is saved to: %APPDATA%\Pangaia Software\MAITO\Models

Step 4: Set as Default

After download completes:

• Click Set as Default to use this model
• Or keep browsing to download more models

Step 5: Optional Benchmark

Run a benchmark to test performance:

• Click Run Benchmark
• Wait 30 seconds for results
• See tokens/second performance

Verifying Configuration

To verify local translation is working:

1. Open MAITO main window
2. Go to Text Translation tab
3. Look for engine indicator showing “Local”
4. Type a test phrase (e.g., “Hello, world!”)
5. Click Translate
6. You should see:
   • Translation appears gradually (streaming)
   • No cost information (unlike DeepL)

Troubleshooting

”Download Failed” Error

Solutions:

• Check your internet connection
• Verify firewall isn’t blocking MAITO
• Retry the download
• Manually download from Settings later

”Insufficient Memory” Error

Solutions:

• Close other applications to free RAM
• Restart your computer
• Choose a smaller model
• Consider adding more RAM to your system

”Model Loading Failed” Error

Solutions:

• Verify model downloaded completely
• Check disk space (models need 2-8GB)
• Try downloading the model again
• Contact support if issue persists

Very Slow Performance

Solutions:

• Your system may not meet minimum requirements
• Close background applications
• Try the CPU-optimized model
• Consider using DeepL instead for better experience

What’s Next?

Now that local translation is configured:

• Verify It’s Truly Offline - Confirm no data is sent
• Download More Models - Try different models
• Benchmark Performance - Compare model speeds
• Manage Disk Space - Remove unused models
• Understanding Local Translation - Learn how it works
• Switch Between Engines - Use both DeepL and Local

Need Help?

If you encounter issues configuring local translation:

• Check our Troubleshooting Guide
• Review Why Is Local Slow
• Contact support: contact@pangaia.software