LM Studio is the easiest way to run large language models privately on your own hardware. It provides a polished desktop interface for downloading, loading, and chatting with any GGUF or safetensors model from Hugging Face — no command line required. Version 0.4.11 (April 2026) adds headless server deployment, an OpenAI-compatible API, MCP client support, and a JavaScript and Python SDK. Everything runs locally: your conversations never leave your machine.
Get more guides like this in your inbox
No spam. Unsubscribe anytime.
System Requirements
Before downloading, confirm your hardware meets the minimum requirements. LM Studio supports Apple Silicon Macs, Windows x64/ARM, and Linux x64/ARM64.
| Platform | Minimum | Recommended |
|---|---|---|
| macOS (Apple Silicon) | M1, macOS 14.0 (Sonoma), 8GB RAM | M2/M3/M4, 16GB+ RAM |
| Windows x64 | AVX2 CPU, 16GB RAM, 4GB VRAM | RTX 4090 or higher, 32GB RAM |
| Windows ARM | Snapdragon X Elite | 16GB+ RAM |
| Linux x64/ARM64 | Ubuntu 20.04+, AppImage | Ubuntu 22.04, 16GB+ RAM, NVIDIA GPU |
Step 1 — Download and Install LM Studio
Visit lmstudio.ai/download and download the installer for your operating system. The current version is 0.4.11.
Step 2 — Choose and Download a Model
Open the Discover tab (or press Cmd+2 on Mac / Ctrl+2 on Windows/Linux). Search for a model family by name — try 'Qwen', 'Llama', 'Gemma', or 'Mistral'. LM Studio pulls directly from Hugging Face, so any GGUF model is available.
Understanding Quantization (Q Levels)
When you search for a model, you will see multiple download options named things like Q3_K_S, Q4_K_M, Q8_0. These are quantized versions of the same model — compressed to different degrees to trade file size against quality.
| Quantization | File Size (7B model) | Quality | Best For |
|---|---|---|---|
| Q3_K_S | ~2.0 GB | Lower | Very limited RAM (8GB Mac) |
| Q4_K_M | ~2.7 GB | Good | Most users — best balance |
| Q5_K_M | ~3.2 GB | Better | 16GB+ RAM, quality priority |
| Q8_0 | ~4.7 GB | Near-lossless | 24GB+ VRAM, maximum quality |
Step 3 — Load the Model
Once the download completes, open the Chat tab. Click the model loader at the top of the screen (or press Cmd+L on Mac / Ctrl+L on Windows/Linux). Select the model you downloaded from the list. Optionally adjust load parameters such as context length and GPU offload layers.
Step 4 — Your First Conversation
Once the model is loaded, type a message in the chat box and press Enter. The model will respond locally — no internet connection required, no data sent to any server. You can adjust system prompt, temperature, and context window in the settings panel on the right.
Step 5 — Enable the Local API Server
LM Studio includes an OpenAI-compatible local API server. Enable it from the Developer tab (or the server icon in the sidebar). Once running, any application that supports the OpenAI API can connect to it at http://localhost:1234/v1 — including Open WebUI, Cursor, Continue.dev, and your own scripts.
# Connect to LM Studio from Python (using the OpenAI SDK)
from openai import OpenAI
client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")
completion = client.chat.completions.create(
model="qwen3.5-4b", # use the model name shown in LM Studio
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "What is quantization in the context of LLMs?"}
]
)
print(completion.choices[0].message.content)Recommended First Models by Hardware
Not sure which model to start with? Use this table as a starting point based on your available RAM or VRAM.
| Hardware | Recommended Model | Quantization | Why |
|---|---|---|---|
| Mac mini M4 (16GB) | Qwen3.5 4B or Gemma 3 4B | Q8_0 | Fits in unified memory, fast on Apple Silicon |
| Mac mini M4 Pro (24GB) | Qwen3.5 8B or Llama 3.1 8B | Q8_0 | Full quality 8B fits comfortably |
| Mac Studio M4 Max (36GB) | Qwen3.5 14B or Mistral Small 4 24B | Q6_K | Excellent quality, near-lossless |
| RTX 4090 24GB | Llama 3.1 8B or Mistral Small 4 24B | Q8_0 | Full VRAM utilisation, fastest inference |
| RTX 4070 12GB | Qwen3.5 7B or Llama 3.1 8B | Q4_K_M | Fits in VRAM with headroom |
| 8GB RAM (any) | Qwen3.5 1.7B or Gemma 3 1B | Q4_K_M | Only option at this memory level |
Privacy Notes
LM Studio is designed for private, local inference. By default, no model data, conversation history, or prompts are sent to any external server. The Discover tab fetches model metadata from Hugging Face to populate search results, but the actual model weights are downloaded directly to your machine. The local API server only listens on localhost (127.0.0.1) by default — it is not exposed to your network unless you explicitly change the bind address.
Next Steps
Once you are comfortable with basic chat, explore these features: Chat with Documents (RAG) lets you load PDFs and text files and ask questions about them. MCP client support connects LM Studio to external tools like web search and file access. The CLI (lms command) lets you load models and start the server from a terminal. The headless deployment option (llmster) runs LM Studio on a Linux server with no GUI — useful for a home server or NAS.