How to Set Up LM Studio: Complete Local LLM Guide (2026)

Install LM Studio on Windows, Mac, or Linux

LM Studio ships as a single installer per platform. No dependencies or terminal use are required for this step.

Install on Windows

Download the Windows build from lmstudio.ai/download. The page detects your system and serves the x64 AVX2 installer, roughly 500 MB.

Run the `.exe` file, accept the default install location, and let it finish. LM Studio adds itself to the Start Menu automatically.

Install on macOS

Download the `.dmg` file from the same download page. Apple Silicon and Intel builds are both available, and the page auto-detects which one your Mac needs.

Open the `.dmg`, drag LM Studio into Applications, then launch it from Launchpad. macOS asks you to confirm opening an app downloaded from the internet the first time.

Install on Linux

Linux ships as an `.AppImage`, available in x64 and ARM64 builds.

chmod +x LM-Studio-0.4.20-x64.AppImage
./LM-Studio-0.4.20-x64.AppImage

No package manager install is required. The AppImage runs directly once it has execute permission.

On first launch, LM Studio opens to a search screen for finding models. No account or sign-in is needed to use the app locally.

Download and Load Your First Model

LM Studio pulls models directly from Hugging Face in GGUF format, the file format used by the llama.cpp inference engine that powers the app.

Search for a Model

Click the search icon in the left sidebar (or press Cmd/Ctrl+Shift+M) to open the model browser. Type a model name, for example `llama-3.3-8b`, and LM Studio lists the available GGUF builds with a compatibility badge showing whether your hardware can run each one.

Recommended First Models

Q4_K_M is the quantisation LM Studio recommends by default for most systems. It keeps output quality close to the full-precision model while cutting file size by roughly 4x.

Download and Load

Click Download next to your chosen model. LM Studio shows a progress bar with download speed and time remaining. A 4.9 GB model takes 3-5 minutes on a 200 Mbps connection.

Once downloaded, click Load Model, or pick it from the model dropdown at the top of the chat window. LM Studio reads the model into RAM (and VRAM, if a compatible GPU is detected) and reports a Ready state within 10-30 seconds depending on model size and disk speed.

Chat With a Local Model

With a model loaded, the chat interface behaves like ChatGPT's, entirely offline.

Basic Chat

Type a message in the input box and press Enter. LM Studio streams the response token by token, and the status bar at the bottom of the window shows tokens per second, a rough measure of how fast your hardware is generating text.

Set a System Prompt

Click the settings icon in the chat sidebar to open the system prompt field. This sets persistent instructions the model follows for the rest of the conversation, for example: `You are a concise coding assistant. Answer in code blocks with minimal explanation.`

Adjust Context Length

Models load with a default context window read from the GGUF file's own metadata, often only 4,096 or 8,192 tokens. For longer conversations or large pasted documents, raise this in the model load settings before loading, or reload the model after changing it.

Start the LM Studio Local API Server

The Local Server turns LM Studio into an OpenAI-compatible API endpoint, the feature that lets other tools such as Claude Code, custom scripts, LangChain, or n8n talk to your local model.

Start the Server from the GUI

Open the Developer tab in the left sidebar (the `` icon), select a loaded model, and toggle Start Server. The server binds to `http://localhost:1234` by default.

Start the Server from the Command Line

LM Studio ships an `lms` CLI alongside the app, useful for scripting or running the server without opening the GUI.

lms server start --port 1234

Expected output:

Server started on port 1234
Serving on http://localhost:1234

Test the Server

curl http://localhost:1234/v1/models

Expected output, trimmed:

{
  "data": [
    {
      "id": "llama-3.3-8b-instruct",
      "object": "model",
      "owned_by": "organization-owner"
    }
  ],
  "object": "list"
}

If this returns your loaded model's ID, the server is reachable. Any OpenAI SDK, in Python, JavaScript, or plain curl, can now point its base URL at `http://localhost:1234/v1` instead of `api.openai.com` and reuse the exact same request format.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="not-needed")

response = client.chat.completions.create(
    model="llama-3.3-8b-instruct",
    messages=[{"role": "user", "content": "Explain what a REST API is in one sentence."}]
)
print(response.choices[0].message.content)

Set Up an LM Studio API Key

By default, the local server accepts requests from anyone who can reach `localhost:1234`. That is fine for solo use on your own machine and risky the moment you expose the port to a network.

Enable Authentication

In the Developer tab, open Server Settings and toggle Require Authentication. This feature requires LM Studio 0.4.0 or later.

Generate an API Key

Click Manage Tokens inside Server Settings, then Create Token. Name the token, for example `claude-code` or `n8n-workflow`, set its permissions, and copy the value shown. LM Studio only displays the token once.

Use the Key

Send the token as a Bearer header on every request:

curl http://localhost:1234/v1/models \
  -H "Authorization: Bearer YOUR_TOKEN_HERE"

Requests without a valid token now return a 401 once authentication is switched on.

Enable Plugins and Web Search in LM Studio

Plugins are LM Studio's system for giving a local model tools it does not have by default: filesystem access, shell commands, or in this case, live web search. The plugin system has been in beta since the 0.4.x release line.

Install a Web Search Plugin

Open the Plugins panel from the sidebar and search the LM Studio Hub for `web-search`. Several community plugins are listed; a DuckDuckGo-backed option works with no API key needed, which makes it the simplest starting point.

Click Install, then enable the plugin for your current chat.

Use Web Search in a Chat

With the plugin active and a tool-calling-capable model loaded (Llama 3.3, Qwen2.5, and Gemma 4 all support tool calls), ask a question that needs current information:

What is the latest stable release version of LM Studio?

The model calls the search tool on its own. LM Studio shows a small expandable card in the chat with the query it ran and the pages it read before answering.

Connect LM Studio to Claude Code

LM Studio exposes an Anthropic-compatible endpoint alongside the OpenAI-compatible one, which means Claude Code can talk to a local model with two environment variables and no proxy layer in between.

Step 1: Start the Server

lms server start --port 1234

Step 2: Set Environment Variables

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio

Add both lines to `~/.zshrc` or `~/.bashrc` to keep them across terminal sessions. If you turned on authentication in the previous section, use your real API token in place of the placeholder `lmstudio` value.

Step 3: Run Claude Code Against the Local Model

claude --model openai/gpt-oss-20b

Replace the model name with whatever you have loaded in LM Studio. If Claude Code reports it cannot find the model, confirm the exact ID first:

curl http://localhost:1234/v1/models

Use the `id` field from that response exactly. A mismatched model name is the most common connection failure.

VS Code Configuration

For the Claude Code VS Code extension, add the same variables to `settings.json`:

"claudeCode.environmentVariables": [
  { "name": "ANTHROPIC_BASE_URL", "value": "http://localhost:1234" },
  { "name": "ANTHROPIC_AUTH_TOKEN", "value": "lmstudio" }
]

GPU Offloading and Performance Settings

GPU offload is the single setting with the biggest effect on generation speed.

How GPU Offload Works

Each model layer can run on GPU (fast) or CPU (slow). The GPU Offload slider in the model load settings controls how many layers move to VRAM. Setting it to the model's full layer count, which LM Studio displays, for example 32/32 for many 7B models, runs the entire model on GPU when VRAM allows it.

VRAM Budgeting

A rough rule: a Q4_K_M quantised model needs VRAM close to its file size, plus 1-2 GB overhead for context and the KV cache. An 8 GB model with a full 32,768-token context comfortably fits an RTX 4090's 24 GB VRAM, with room to run a second, smaller model alongside it.

When a model does not fit locally, for example a 70B model needing 40 GB or more VRAM, renting a GPU by the hour costs less than buying hardware for occasional large-model use. An A100 80GB on Vast.ai runs 70B-class models at full offload for roughly $1-2 an hour, billed only while the instance is running.

Apple Silicon

On Apple Silicon Macs, LM Studio uses Metal for GPU acceleration automatically. Unified memory means there is no separate VRAM figure to budget for. The GPU offload slider still applies, it draws from the same RAM pool as the rest of the system.

Troubleshooting

Frequently Asked Questions