LM Studio

LM Studio is a proprietary desktop application developed by Element Labs that makes it easy to discover, download, and run large language models locally on your own computer. It provides a graphical interface for browsing the Hugging Face Hub, downloading GGUF models, and running them with GPU acceleration — no command-line skills required.

The app includes a built-in OpenAI-compatible local inference server, allowing any tool or application that supports OpenAI's API to connect to locally running models. It supports NVIDIA CUDA, AMD ROCm, Apple Silicon Metal, Vulkan, and CPU-only inference. LM Studio is available at lmstudio.ai for macOS, Windows, and Linux.

Download the appropriate installer for your platform from lmstudio.ai:

Platform	Download
macOS (Apple Silicon)	LM Studio--arm64.dmg
macOS (Intel)	LM Studio--x64.dmg
Windows	LM Studio--x64.exe
Linux	LM Studio--x86_64.AppImage

No Python or Node.js dependencies required. The app bundles everything needed for local LLM inference.

LM Studio includes a built-in model browser that connects to the Hugging Face Hub, allowing you to search, filter, and download GGUF models directly from the app. You can browse by model name, parameter count, quantization type, or popularity. The search bar supports natural language queries like "7B Q4_K_M" to find specific quantized variants.

Key Features

• One-click download — Browse and download models with a single click
• Automatic quantization detection — Shows available GGUF quantization variants
• Local model library — Browse, search, and manage downloaded models
• Model info panel — View architecture, context length, license, and author details
• Multi-model loading — Keep multiple models loaded and switch between them
• Custom model paths — Import models from external folders

LM Studio provides a built-in HTTP server that exposes an OpenAI-compatible API, enabling any application that supports OpenAI's chat completions endpoint to use locally running models. This allows you to use LM Studio as a drop-in replacement for OpenAI with tools like Cursor, Continue.dev, Open Interpreter, and custom scripts.

Server Details

• Default endpoint: http://localhost:1234/v1
• API format: OpenAI-compatible chat completions, embeddings, and models
• Cross-Origin: CORS enabled by default for web app integration
• Streaming: Full server-sent events (SSE) support
• Multi-user: Basic API key authentication support

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "model-identifier",
    "messages": [{"role": "user", "content": "Hello!"}],
    "temperature": 0.7
  }'

LM Studio's chat interface provides a polished conversational experience with multi-turn对话, system prompt configuration, and model parameter tuning. The UI is designed to be accessible to non-technical users while still offering advanced controls for power users.

• Multi-turn chat — Full conversation history with context window management
• Parameter controls — Adjust temperature, top-p, max tokens, frequency/presence penalties
• System prompts — Custom system messages for behavior control
• Model comparison — Side-by-side chat to compare model outputs
• Local RAG — Attach files and documents for in-context retrieval
• Context viewer — See what the model sees with token-level context inspection
• Multi-modal support — Image inputs for vision-capable models

LM Studio supports hardware acceleration across all major GPU vendors. On Apple Silicon Macs, it uses Metal for near-native performance. On Windows and Linux, it supports NVIDIA CUDA and AMD ROCm for GPU offloading. The app automatically detects your hardware and applies optimal settings, but you can manually configure GPU layers, thread count, and context size.

Setting	Description
GPU Offload	Select number of layers to offload to GPU
Context Length	Maximum tokens the model can process at once
Thread Count	CPU threads for prompt evaluation
Flash Attention	Memory-efficient attention for longer contexts
KV Cache Quantization	Reduce memory usage with Q8/Q4 key-value cache