LM Studio: Complete Guide to Local LLM Inference (2026)
Run powerful AI models on your own hardware with zero cloud dependency
Hypereal AI TeamHyperealで構築を始めよう
Kling、Flux、Sora、Veoなどに単一のAPIでアクセス。無料クレジットで開始、数百万規模まで拡張可能。
クレジットカード不要 • 10万人以上の開発者 • エンタープライズ対応
LM Studio: Complete Guide to Local LLM Inference (2026)
LM Studio is a desktop application that lets you download, run, and interact with large language models entirely on your local hardware. No cloud dependency, no API keys, no usage fees, and complete privacy. Your data never leaves your machine.
In 2026, local LLM inference has become surprisingly practical. With optimized quantization formats like GGUF, even consumer hardware can run models that rival cloud APIs for many tasks. This guide covers everything you need to know about LM Studio: installation, model selection, configuration, performance optimization, and API setup.
What Is LM Studio?
LM Studio is a free desktop application for macOS, Windows, and Linux that provides:
- A model discovery and download interface (browsing Hugging Face)
- A chat UI for interacting with models
- An OpenAI-compatible local API server
- Model management (download, delete, organize)
- Configurable inference parameters (temperature, context length, GPU layers)
- Support for GGUF, MLX, and other quantized model formats
Why Run Models Locally?
| Advantage | Details |
|---|---|
| Privacy | Data never leaves your machine |
| No cost | No API fees or subscriptions |
| No rate limits | Use as much as you want |
| Offline | Works without internet after model download |
| Customization | Full control over parameters and system prompts |
| Speed | No network latency (GPU inference can be very fast) |
System Requirements
LM Studio runs on a wide range of hardware, but performance scales significantly with GPU memory and system RAM.
Minimum Requirements
| Component | Minimum | Recommended |
|---|---|---|
| OS | macOS 13+, Windows 10+, Ubuntu 22.04+ | Latest version |
| RAM | 8 GB | 16-32 GB |
| GPU | Not required (CPU mode) | 8+ GB VRAM |
| Storage | 10 GB free | 50+ GB free |
| CPU | Any 64-bit | Apple Silicon or modern x86 |
GPU Compatibility
| GPU Type | Support | Notes |
|---|---|---|
| NVIDIA (CUDA) | Full | Best performance on Windows/Linux |
| Apple Silicon (Metal) | Full | Excellent performance on macOS |
| AMD (ROCm/Vulkan) | Partial | Linux ROCm works well, Vulkan on Windows |
| Intel Arc | Partial | Improving support via Vulkan |
| CPU only | Yes | Slow but functional for small models |
Step 1: Install LM Studio
macOS
# Download from the website
# Visit https://lmstudio.ai and download the .dmg file
# Or install via Homebrew
brew install --cask lm-studio
Windows
Download the installer from lmstudio.ai and run it. LM Studio installs to your user directory and does not require administrator privileges.
Linux
# Download the AppImage from lmstudio.ai
chmod +x LM-Studio-*.AppImage
./LM-Studio-*.AppImage
# Or use Flatpak (if available)
flatpak install flathub ai.lmstudio.LMStudio
Step 2: Download Your First Model
After launching LM Studio, use the Discover tab to browse and download models.
Recommended Models by Hardware (2026)
| Hardware | Model | Size | Quality |
|---|---|---|---|
| 8 GB RAM (CPU) | Qwen 3 0.6B Q8 | 0.8 GB | Basic tasks |
| 16 GB RAM (CPU) | Llama 4 Scout 8B Q4_K_M | 5 GB | Good for chat |
| 8 GB VRAM | Qwen 3 14B Q4_K_M | 9 GB | Very good |
| 12 GB VRAM | Qwen 3 32B Q4_K_M | 19 GB | Excellent |
| 16 GB VRAM | Llama 4 Scout 109B Q3_K_M | 14 GB | Excellent |
| 24 GB VRAM (RTX 4090) | DeepSeek Coder V3 Q4_K_M | 18 GB | Near-cloud quality |
| Apple M4 Pro 24GB | Qwen 3 32B Q4_K_M | 19 GB | Excellent |
| Apple M4 Max 64GB | Llama 4 Maverick Q4_K_M | 55 GB | Cloud-competitive |
How to Download a Model
- Go to the Discover tab in LM Studio
- Search for the model name (e.g., "Qwen 3 14B")
- Select the GGUF quantization you want (Q4_K_M is a good default)
- Click Download
- Wait for the download to complete (models are 2-60+ GB)
Understanding Quantization
Quantization reduces model size and memory usage at the cost of some quality. Here is a guide to common GGUF quantization levels:
| Quantization | Bits | Size vs. FP16 | Quality Impact |
|---|---|---|---|
| Q2_K | 2-bit | ~25% | Significant quality loss |
| Q3_K_M | 3-bit | ~35% | Noticeable quality loss |
| Q4_K_M | 4-bit | ~45% | Minimal quality loss (recommended) |
| Q5_K_M | 5-bit | ~55% | Very minor quality loss |
| Q6_K | 6-bit | ~65% | Near-lossless |
| Q8_0 | 8-bit | ~85% | Effectively lossless |
| FP16 | 16-bit | 100% | Original quality |
Q4_K_M is the sweet spot for most users: minimal quality degradation with roughly half the memory usage of the full model.
Step 3: Chat with Your Model
- Go to the Chat tab
- Select your downloaded model from the dropdown
- Start typing messages
Useful Chat Settings
| Setting | Default | Recommended | Purpose |
|---|---|---|---|
| Temperature | 0.7 | 0.1-0.3 for code, 0.7-0.9 for creative | Controls randomness |
| Context Length | 4096 | Max your hardware supports | How much text the model remembers |
| GPU Layers | Auto | All (if VRAM allows) | How many layers run on GPU |
| System Prompt | None | Set per use case | Instructs the model's behavior |
Example System Prompts
For coding assistance:
You are an expert software developer. Write clean, well-documented code.
Always include error handling and type annotations. Prefer standard library
solutions over third-party dependencies. Explain your reasoning briefly.
For writing assistance:
You are a professional editor. Help improve writing clarity, grammar, and
structure. Suggest specific edits rather than general advice. Maintain the
author's voice and intent.
Step 4: Use the Local API Server
LM Studio includes an OpenAI-compatible API server. This lets you use local models with any tool that supports the OpenAI API format -- including Cursor, Continue, Cline, Aider, and custom applications.
Start the API Server
- Go to the Developer tab (or Local Server tab)
- Select your model
- Click Start Server
- The server runs at
http://localhost:1234by default
Test the API
curl http://localhost:1234/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-14b",
"messages": [
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a Python function to flatten a nested dictionary."}
],
"temperature": 0.2,
"max_tokens": 1000
}'
Use with Python
from openai import OpenAI
# Point to LM Studio's local server
client = OpenAI(
base_url="http://localhost:1234/v1",
api_key="not-needed" # LM Studio doesn't require an API key
)
response = client.chat.completions.create(
model="qwen3-14b",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Explain how HTTP caching works."}
],
temperature=0.3
)
print(response.choices[0].message.content)
Connect to Cursor
- Open Cursor > Settings > Models
- Add a custom model:
- API Key:
lm-studio(any non-empty string) - Base URL:
http://localhost:1234/v1 - Model name: The name of your loaded model
- API Key:
- Select the model in Cursor's chat or agent panel
Connect to Continue (VS Code)
// ~/.continue/config.json
{
"models": [
{
"title": "LM Studio - Qwen 3 14B",
"provider": "openai",
"model": "qwen3-14b",
"apiBase": "http://localhost:1234/v1",
"apiKey": "not-needed"
}
]
}
Connect to Aider
# Use LM Studio as the backend for Aider
aider --model openai/qwen3-14b \
--openai-api-base http://localhost:1234/v1 \
--openai-api-key not-needed
Step 5: Optimize Performance
Maximize GPU Offloading
The most impactful performance setting is GPU offloading. Set GPU layers to the maximum your VRAM allows:
| Model Size | GPU VRAM Needed (Q4_K_M) | Approximate Speed |
|---|---|---|
| 7-8B | 5-6 GB | 30-60 tokens/sec |
| 14B | 9-10 GB | 20-40 tokens/sec |
| 32B | 19-22 GB | 10-25 tokens/sec |
| 70B | 40-45 GB | 5-15 tokens/sec |
Context Length vs. Speed
Longer context windows use more memory and slow down inference. Set context length based on your actual needs:
General chat: 4096-8192 tokens
Code assistance: 8192-16384 tokens
Document analysis: 16384-32768 tokens
Large codebase: 32768-65536 tokens
Memory Tips
- Close other applications to free RAM for model loading
- Use Q4_K_M quantization as the default (best quality/size ratio)
- If a model barely fits in VRAM, try Q3_K_M to free some memory
- On Apple Silicon, unified memory means the system RAM is shared between CPU and GPU. A 32 GB Mac can fully load models that need 28-30 GB
LM Studio vs. Ollama
LM Studio and Ollama are the two most popular local inference tools. Here is how they compare:
| Feature | LM Studio | Ollama |
|---|---|---|
| Interface | GUI + API | CLI + API |
| Model format | GGUF, MLX | GGUF (via Modelfile) |
| Model discovery | Built-in browser | ollama pull |
| API compatibility | OpenAI-compatible | OpenAI-compatible |
| Platform | macOS, Windows, Linux | macOS, Windows, Linux |
| Resource usage | Higher (Electron app) | Lower (CLI) |
| Ease of use | Easier for beginners | Easier for CLI users |
| Price | Free | Free |
Choose LM Studio if you prefer a graphical interface for browsing, downloading, and managing models. Choose Ollama if you prefer a CLI-first workflow and want lower resource overhead.
Frequently Asked Questions
Is LM Studio free? Yes, LM Studio is completely free for personal use. There are no API fees, subscriptions, or usage limits.
What models should I start with? Start with Qwen 3 14B Q4_K_M if you have 16 GB RAM or 8+ GB VRAM. For coding specifically, try DeepSeek Coder V3 or Qwen 2.5 Coder.
Can local models match cloud API quality? For many tasks, yes. A well-quantized 32B or 70B parameter model running locally produces output comparable to GPT-4o for coding, writing, and analysis. For the most demanding tasks, cloud models (GPT-5, Claude Opus 4) still have an edge.
Can I use LM Studio with Cursor/Cline/Aider? Yes. LM Studio's OpenAI-compatible API server works with any tool that supports custom OpenAI endpoints. See the configuration examples in Step 4.
Does LM Studio work offline? Yes. After downloading a model, LM Studio works completely offline. No internet connection is needed for inference.
How much disk space do I need? Models range from 1 GB (small 3B models) to 60+ GB (large 70B+ models). Plan for 10-50 GB depending on how many models you want to keep downloaded.
Wrapping Up
LM Studio makes local LLM inference accessible to everyone. With the right model for your hardware, you get a private, free, offline AI assistant that handles coding, writing, analysis, and creative tasks. The OpenAI-compatible API server means your local models integrate seamlessly with Cursor, Aider, Continue, and custom applications.
For tasks that require cloud-level AI capabilities -- like AI-generated images, video, and audio -- try Hypereal AI free -- 35 credits, no credit card required. Combine LM Studio's local text generation with Hypereal's cloud API for media generation to build powerful AI applications while keeping your costs low.
関連記事
Download Hypereal Agent
Run a local AI media workspace for image generation, video prompts, model selection, credit tracking, and saved artifacts.
