LM Studio Setup Guide for Phase 2 Testing

Why LM Studio?

• ✅ FREE - No API costs
• ✅ Fast - Runs locally on your GPU
• ✅ Private - Your code never leaves your machine
• ✅ DeepSeek-R1 - Strong reasoning for code refactoring

Step 1: Download LM Studio

1. Go to https://lmstudio.ai/
2. Download for Windows
3. Install and launch

Step 2: Download DeepSeek-R1 Model

In LM Studio:

1. Click “Search” tab
2. Search for: deepseek-r1-distill-qwen-32b
3. Recommended model: bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF
   • Download the Q4_K_M quantization (~20GB)
   • Good balance of speed and quality
   • Fits in 16GB RAM

Alternative (smaller):

• DeepSeek-R1-Distill-Qwen-14B - Faster, requires ~10GB RAM

Step 3: Load Model

1. Go to “Chat” tab
2. Click “Select a model”
3. Choose the downloaded DeepSeek-R1 model
4. Wait for it to load (shows “Model loaded” in bottom right)

Step 4: Start Local Server

1. Click “Local Server” tab (left sidebar)
2. Click “Start Server” button
3. Default endpoint: http://localhost:1234/v1
4. Keep LM Studio running!

Step 5: Test with Ouroboros

Open PowerShell in your project:

# Activate virtual environment
cd 'g:\Just a Idea'
.\venv\Scripts\Activate.ps1
$env:PYTHONPATH='g:\Just a Idea'

# Test connection
python scripts/generate_refactor_plan.py test-connection --provider lmstudio

# Generate refactor plan (FREE!)
python scripts/generate_refactor_plan.py generate "Rename function foo to bar" --file src/utils.py --provider lmstudio

Configuration Options

Environment Variables (Optional)

# Change LM Studio endpoint (if using different port)
$env:LMSTUDIO_BASE_URL = "http://localhost:1234/v1"

# Set as default provider
$env:REASONER_PROVIDER = "lmstudio"

Update Model Name

If you downloaded a different model, update src/reasoner/config.py:

LMSTUDIO_CONFIG = ModelConfig(
    model_name="your-model-name-here",  # Match LM Studio
    max_tokens=8192,
    context_window=32_000,
    ...
)

Troubleshooting

Error: “Connection refused”

• ✅ Make sure LM Studio Local Server is running
• ✅ Check the port is 1234 (default)
• ✅ Verify model is loaded in Chat tab

Error: “Model not found”

• ✅ Update model_name in config.py to match LM Studio exactly
• ✅ Check model name in LM Studio’s loaded models list

Slow responses

• ✅ Use smaller quantization (Q4_K_M or Q4_K_S)
• ✅ Try 14B model instead of 32B
• ✅ Enable GPU acceleration in LM Studio settings

Performance Tips

1. GPU Acceleration:
   • LM Studio Settings → Enable GPU offloading
   • Set layers to offload = all available
2. Context Length:
   • Reduce context_window in config if model is slow
   • Start with 16K tokens for faster responses
3. Quantization:
   • Q4_K_M: Best balance (recommended)
   • Q4_K_S: Faster, slightly lower quality
   • Q5_K_M: Higher quality, slower

Comparing Providers

Provider	Cost	Speed	Quality	Context
LM Studio (DeepSeek-R1)	FREE	Medium	High	32K
Claude 3.5 Sonnet	$0.06/req	Fast	Highest	200K
OpenAI GPT-4o	$0.025/req	Fast	High	128K
Jamba 1.5 Mini	$0.001/req	Fast	Medium	256K

Next Steps

Once LM Studio is working:

1. Test with simple refactors: “Rename function X to Y”
2. Try complex tasks: “Extract validation logic into separate function”
3. Compare quality with mock provider
4. When ready, add OpenAI credits for production use

---

Questions? LM Studio has excellent docs at https://lmstudio.ai/docs