LM Studio for Beginners: Run a Chat Model Offline on Mac, Windows or Linux
Note: Some future links on this site may be affiliate links. We never claim hands-on testing; guides are researched from the public sources listed in each article.
How we wrote this: A how-to based on LM Studio's official documentation (linked below). We have not benchmarked any hardware or model. The app changes often, so check the official docs before relying on a specific menu or shortcut.
If a command line puts you off, LM Studio is one of the friendlier ways to run an AI model on your own computer. It's a desktop app with a chat window, a built-in model downloader and an optional local server. This guide covers what you need, how to pick a download, and what does and doesn't stay on your machine.
Check your computer first
LM Studio's system requirements at the time of writing:
| System | What the docs list |
|---|---|
| Mac | Apple Silicon (M1 to M4) and macOS 14.0 or newer. 16GB+ RAM recommended; 8GB Macs can work with smaller models and modest context sizes. Intel Macs aren't supported. |
| Windows | x64 or ARM (Snapdragon X Elite). x64 CPUs need AVX2. At least 16GB RAM and at least 4GB of dedicated VRAM recommended. |
| Linux | x64 or ARM64, distributed as an AppImage. Ubuntu 20.04 or newer. |
Memory is the real limit. A model has to fit in RAM (or VRAM) with room left for the conversation. Private AI Computer's memory calculator estimates whether a given model and quantization fit on a given machine.
Download a model
Per the getting-started docs, LM Studio runs models with llama.cpp (GGUF files) on Mac, Windows and Linux, and with Apple's MLX on Apple Silicon Macs. Its model downloader searches Hugging Face by keyword, by user/model name, or by a pasted Hugging Face URL.
You'll usually see several download options for one model, with names like Q3_K_S or Q8_0. The docs explain that these are copies of the same model at different degrees of fidelity. "Q" stands for quantization, which compresses the file in exchange for some quality. The docs' advice: choose a 4-bit option or higher if your machine can run it. Our guide to GGUF files and quantization labels explains the names in more detail.
What works offline
LM Studio's offline operation page is the part privacy-minded users should read. According to the docs:
- Chatting with a downloaded model doesn't need the internet, and nothing you type into a chat leaves your device.
- Chatting with documents (drag-and-drop RAG) is processed locally, and the document stays on your machine.
- Running the local server doesn't need the internet.
These do need a connection: searching for models (it calls huggingface.co), downloading models, loading the Discover tab's catalog stats, downloading runtimes, and checking for app updates.
A practical routine: download the models you want while online, then work offline when you handle sensitive material. For a fuller picture of what "private" covers, see What "private AI" really means on Private AI Computer.
The local server
LM Studio can serve a loaded model to other apps. The server docs say to open the Developer tab and switch on "Start server", or run lms server start from a terminal. It offers its own REST API plus OpenAI-compatible and Anthropic-compatible endpoints, so many tools that expect a cloud API can point at your own computer instead.
The server can listen on localhost only or on your local network. Leave it on localhost unless you have a reason to share it, especially on a shared or public network.
Tips for a smooth first hour
- Start small. A small model that loads quickly is a better first test than a large one that barely fits.
- Watch memory. If the computer slows to a crawl, try a smaller model, a lower quantization or a shorter context. Private AI Computer explains how context length uses memory.
- Check the license. Free to download doesn't always mean free for every use. See our guide to open-weight model licenses.
- Compare tools. If you'd rather script things, our Ollama walkthrough covers a command-line route.