LM Studio Installation and Usage Guide: GUI Local LLM Management

What is LM Studio?

LM Studio is a desktop GUI application that lets you download, manage, and run LLMs locally, without a command line. It supports models in GGUF and MLX formats.

Graphical interface: No terminal operations required; run models with a click
Built-in model download: Search and download directly from HuggingFace
OpenAI-compatible API: Automatically provides a localhost:1234/v1 endpoint after startup
Supports Apple Silicon GPU: Native acceleration for M-series chips

Installation

# Official Website Download
# https://lmstudio.ai/

# Or Homebrew
brew install --cask lm-studio

Model Download and Management

Steps

Open LM Studio → click the Search icon on the left
Search for a model, for example qwen2.5-coder-7b
Select the GGUF format (Q4_K_M quantization is the most balanced)
Click Download

Recommended Models

Model	Size	Recommended Quantization	Use Case
Qwen 2.5 Coder 7B	4.7GB	Q4_K_M	Code
Qwen 2.5 32B	19GB	Q4_K_M	General
DeepSeek Coder V2	15GB	Q4_K_M	Code
Llama 3.1 8B	4.9GB	Q4_K_M	English
Mistral Nemo 12B	7GB	Q4_K_M	Fast

Running Models

Click the Chat icon on the left.
Select the downloaded model from the top dropdown menu.
Adjust the parameters on the right:
- GPU Offload: Set to Max (load everything onto the GPU)
- Context Length: 8192 (daily use) / 32768 (long documents)
- Temperature: 0.7
Once the model is loaded, you can start chatting.

API Server Mode

Starting the Server

Click Local Server on the left, then Start Server

Default port: http://localhost:1234

Using in Claude Code

{
  "apiKey": "not-needed",
  "baseURL": "http://localhost:1234/v1",
  "model": "qwen2.5-coder-7b-instruct"
}

Calling from Python

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="not-needed"
)

response = client.chat.completions.create(
    model="qwen2.5-coder-7b-instruct",
    messages=[{"role": "user", "content": "Write a Python function to sort a list"}]
)

print(response.choices[0].message.content)

MLX vs GGUF Format Selection

LM Studio supports both GGUF and MLX formats:

Feature	GGUF	MLX
Speed (M3 Ultra)	Fast	Faster
Model library breadth	Very extensive	Less extensive
Quantization options	Q2-Q8, K-quants	Q4, Q8
Cross-platform	✅	Apple Silicon only
Recommended use cases	General purpose	Maximizing performance on M-series chips

Recommendation: When an MLX version is available, prefer MLX; otherwise, use GGUF Q4_K_M.

Advanced Configuration

Running Multiple Models Simultaneously

# Terminal 1: General Model
curl http://localhost:1234/v1/chat/completions -d '{"model":"qwen2.5-32b","messages":[...]}'
# Terminal 2: Code Model (Different Port)
# Start a second server in LM Studio, with the port changed to 1235
curl http://localhost:1235/v1/chat/completions -d '{"model":"qwen2.5-coder-7b","messages":[...]}'

GPU Layer Adjustment

Adjust GPU Offload in the right panel of LM Studio:

Max: Load everything into GPU (fastest, requires sufficient VRAM/unified memory)
Auto: Automatic balancing
Manual: Specify the number of layers (suitable for running multiple models simultaneously)

LM Studio vs oMLX vs Ollama Comparison

Feature	LM Studio	oMLX	Ollama
Interface	GUI	GUI	CLI
Model download	Built-in	HuggingFace	`ollama pull`
Model format	GGUF + MLX	MLX	GGUF
API compatibility	✅	✅	✅
GPU acceleration	✅	✅ (MLX)	✅ (Metal)
Beginner friendliness	⭐⭐⭐⭐⭐	⭐⭐⭐	⭐⭐⭐