LM Studio is the easiest way to run AI models like Qwen, Gemma, Llama and DeepSeek on your own computer, with a ChatGPT-style app and no command line. It is free, works offline after you download a model, and can also act as a local API server for other apps.
This beginner guide walks you through everything: system requirements, installing, picking a model that fits your PC, chatting with documents, the settings that matter, running the local server, the lms command line, MCP tools and fixes for common problems. Details are checked against LM Studio’s official docs as of November 2026.
- What: a free desktop app to download and run open AI models locally on Windows, Mac and Linux.
- Needs: Apple Silicon Mac on macOS 14+, or a Windows PC with AVX2 (x64) or Snapdragon X; 16 GB RAM and 4 GB+ VRAM recommended.
- Get a model: open the Discover tab (Ctrl/Cmd + 2), search, pick a 4-bit quant (Q4_K_M) that fits your memory, download.
- Chat: load the model at the top of the Chat tab, then talk to it or attach documents for offline RAG.
- API: start the server from the Developer tab or
lms server start; default addresshttp://localhost:1234, OpenAI-compatible. - Engines: llama.cpp (GGUF) on all systems, plus MLX on Apple Silicon.
What Is LM Studio?
LM Studio is a desktop app that bundles three things: a model browser connected to Hugging Face, a chat interface, and a local server. Everything runs on your hardware, so your prompts and files stay on your machine.
Under the hood it uses llama.cpp to run GGUF models on Mac, Windows and Linux, and Apple’s MLX engine on Apple Silicon Macs. You can manage these runtimes with Ctrl + Shift + R (Cmd + Shift + R on Mac).
The 0.4 series, which started in January 2026, added parallel requests with continuous batching, a new REST API and a refreshed interface. LM Studio also makes a separate agent app called Bionic, covered further down.
System Requirements
| System | Supported | Recommended | Not supported |
|---|---|---|---|
| macOS | Apple Silicon (M1 to M4 and newer), macOS 14.0+ | 16 GB+ unified memory (8 GB works for small models) | Intel Macs |
| Windows | x64 with AVX2, or ARM (Snapdragon X Elite) | 16 GB+ RAM and 4 GB+ dedicated VRAM | Old CPUs without AVX2 |
| Linux | x64 and ARM64 AppImage, Ubuntu 20.04+ | Same as Windows | Newer than Ubuntu 22 is less tested |
Your GPU’s VRAM decides how big a model runs fast. For a full table of model sizes vs memory, see how much VRAM do I need for AI.
How to Install LM Studio
- Go to lmstudio.ai and open the Downloads page.
- Pick your system: Mac (Apple Silicon), Windows (x64 or ARM) or Linux (AppImage).
- Windows: run the .exe installer and follow the prompts.
- Mac: open the .dmg and drag LM Studio into Applications.
- Linux: make the AppImage executable (
chmod +x LM-Studio-*.AppImage) and run it. - Open LM Studio. The first launch may suggest a starter model; you can skip it.
- Check runtimes: press Ctrl/Cmd + Shift + R and let it update the llama.cpp (and MLX on Mac) engines.
Download Your First Model
- Open the Discover tab (magnifying glass icon, or Ctrl/Cmd + 2).
- Search by name (for example
qwen3orgemma), byuser/model, or paste a full Hugging Face link. - Open a result and look at the download options. Each one is a different quantization of the same model.
- Pick a 4-bit option or higher, such as Q4_K_M, if your machine can run it. LM Studio marks options that are likely too big for your memory.
- Click Download and wait. Files are often 2 to 20 GB.
- Find it later under My Models, where you can also change the models folder.
Not sure what Q4_K_M or Q8_0 means? Our GGUF quantization guide explains which file to download.
Which Model Should You Pick?
| Your memory (VRAM or Mac RAM) | Model size | Good starter picks |
|---|---|---|
| 8 GB | 3B to 8B at Q4 | Qwen3 4B or 8B, Gemma small variants, Llama 3.2 3B |
| 12 to 16 GB | 8B to 14B at Q4/Q5 | Qwen3 14B, Gemma mid-size, DeepSeek R1 distills |
| 24 GB | 27B to 32B at Q4 | Qwen3 32B class, Gemma 27B class |
| 32 GB+ / 64 GB Mac | Larger MoE models | 30B to 70B class at Q4, MoE models that only activate part of the weights |
- For coding: a coder-tuned model in your size range.
- For reasoning: a “thinking” model, but expect slower answers.
- On a Mac: try the MLX version of a model; it is often faster than GGUF on Apple Silicon.
For DeepSeek specifically, see how to run DeepSeek locally.
Chat and Chat With Your Documents
- Open the Chat tab.
- Click the model picker at the top and select your downloaded model.
- Adjust load settings if offered (context length and GPU offload, explained below), then load.
- Type a message. The first reply is slower while the model warms up.
- Attach a file (PDF, Word, text) to a message to ask questions about it. This works fully offline using retrieval (RAG).
- Set a system prompt in the chat settings sidebar to give the model a role or style.
- Start new chats for new topics so old context does not slow things down.
Settings That Actually Matter
| Setting | What it does | Suggested value |
|---|---|---|
| GPU offload | How many model layers go to your GPU | As high as fits; max is fastest |
| Context length | How much text the model remembers | 4K to 8K to start; raise only if needed (uses more memory) |
| Temperature | Randomness | 0.2 to 0.4 for facts and code, 0.7+ for creative |
| Flash attention | Faster, leaner attention on supported GPUs | On if available |
| KV cache quantization | Shrinks context memory | Try Q8 if you need longer context |
| Keep model in memory | Avoids reloads between chats | On if you have spare RAM |
If a model is much slower than expected, it has probably spilled out of VRAM. Lower the context length or pick a smaller quant.
Run LM Studio as a Local API Server
LM Studio can serve your models to other apps using OpenAI-compatible endpoints, locally or across your network.
- Open the Developer tab in LM Studio.
- Start the server with the toggle, or run
lms server start --port 1234in a terminal. - Load a model (or let the server load one on demand).
- Point your app at
http://localhost:1234/v1with any API key value. - Test it:
curl http://localhost:1234/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model": "your-model-id", "messages": [{"role": "user", "content": "Say hi"}]}'
| API | Endpoints |
|---|---|
| OpenAI-compatible | Chat Completions, Completions (legacy), Embeddings, List Models, Responses, structured output and tool use |
| Anthropic-compatible | Messages |
| LM Studio REST API | Chat (/api/v1/chat), list, load, unload and download models, download status |
This is how you connect LM Studio to coding tools. See using local models in Cursor.
The lms Command Line
| Command | What it does |
|---|---|
lms daemon up |
Starts the LM Studio background service (useful on servers) |
lms get qwen3 |
Downloads a model |
lms server start |
Starts the local API server |
lms chat |
Interactive chat in the terminal |
For machines without a screen, LM Studio also offers a headless version (llmster) for servers and CI.
MCP Tools and Bionic
LM Studio can install MCP servers, so local models can use tools like web search, file access or databases. Only add MCP servers you trust, because tools can read and change data. Our LM Studio MCP guide has setup steps.
LM Studio Bionic (launched July 2026) is a separate agent app for open models. It can run alongside LM Studio for local inference, has skills, an in-app terminal and document previews. Version 1.2.0 arrived on November 5, 2026. Beginners should start with plain LM Studio first.
LM Studio vs Ollama
| Feature | LM Studio | Ollama |
|---|---|---|
| Interface | Full desktop app | CLI first, simple app |
| Finding models | Built-in Hugging Face search | Ollama library plus GGUF import |
| Apple MLX | Yes | Limited |
| Custom models | Per-model settings in the app | Modelfiles |
| Best for | Beginners and visual tinkering | Scripts, servers and Docker |
Many people use both. Full comparison: Ollama vs LM Studio vs llama.cpp.
Troubleshooting
| Problem | Fix |
|---|---|
| “Failed to load model” | Model too big: lower GPU offload or context, or download a smaller quant. Full list in our LM Studio not loading fix |
| Very slow replies | Raise GPU offload, lower context, close other GPU apps |
| App will not start on old PC | CPU lacks AVX2; LM Studio needs it on x64 |
| Intel Mac not supported | Use Ollama or llama.cpp instead |
| Model search empty | Check your internet and firewall; try pasting the Hugging Face URL |
| Gibberish or endless output | Update runtimes (Ctrl/Cmd + Shift + R); try another quant upload |
| Other apps cannot connect | Server not started, wrong port, or firewall blocking 1234 |
| Disk filling up | Delete unused models in My Models, or move the models folder to a bigger drive |
My Take
LM Studio is where I send anyone who wants local AI without fuss. Install it, grab a Q4_K_M model that fits your VRAM, and you have a private ChatGPT in 10 minutes.
Once you are comfortable, turn on the local server and plug it into your editor or scripts. That is where local AI gets really useful.
FAQ
Is LM Studio free?
Yes, the app is free to download and use.
Does LM Studio work offline?
Yes. You need the internet to download the app and models, then chats and document Q&A run offline.
Is LM Studio safe and private?
Models run on your computer, so prompts stay local. Only download models from sources you trust and be careful with MCP tools.
Can LM Studio run on an Intel Mac?
No. It needs an Apple Silicon Mac on macOS 14 or later.
What port does the LM Studio server use?
Port 1234 by default, at http://localhost:1234, with OpenAI-compatible routes under /v1.
Where are LM Studio models stored?
In the models folder shown under My Models, where you can also change the location.
LM Studio or Ollama for beginners?
LM Studio, because everything is visual. Ollama is better for scripts, servers and Docker.




