LM Studio quickstart: from zero to first prompt in ten minutes

What you need before starting

Two requirements: 8 GB of RAM minimum, and a reliable internet connection for the model download. Everything else — the application, the model, the inference engine — downloads as part of these five steps.

No programming experience is required. No package manager, no Python environment, no command line (except one file permission command on Linux). LM Studio handles all of that behind its graphical interface. If you are on Windows or macOS, you can proceed directly to step 1 without any prerequisites. Linux users will need to run a single chmod command to make the AppImage executable before launching; step 1 covers that.

Five steps to a working LM Studio session

Steps 1 and 2 take about two minutes combined. Step 3 depends on your download speed. Steps 4 and 5 are under three minutes together.

LM Studio quickstart — minute, action, and tip

Minute	Action	Tip
0–2	Download and install LM Studio for your platform	macOS: drag to Applications. Linux: chmod +x the AppImage before launching.
2–3	Open the Discover tab and search for a 7B instruct model	Look for a green hardware badge — it means LM Studio thinks the model fits your RAM.
3–8	Download a Q4_K_M quantization variant	Q4_K_M balances quality and size well. Expect 4–4.5 GB for a 7B model.
8–9	Load the model using the Load button	If you have a GPU, drag the layer-offload slider right to accelerate inference.
9–10	Open Chat, write a short system prompt, send your first message	Keep the first system prompt simple — "You are a helpful assistant." is enough.

Step 1 — Download and install

Go to the LM Studio download page and pick your platform. On Windows, the installer is a standard .exe wizard. On macOS, open the .dmg and drag the app to Applications. On Linux, save the AppImage to a folder, open a terminal, navigate to that folder, and run chmod +x LMStudio-*.AppImage, then double-click to launch. The application opens to a welcome screen with a left sidebar.

Step 2 — Find a model

Click the Discover icon in the sidebar (it looks like a compass). Type a model name in the search bar — llama-3-8b-instruct or mistral-7b-instruct both work well for a first session. The results show cards for each variant, each labeled with a file size, parameter count, and a color-coded hardware badge. Green means the model is expected to fit your available memory. Yellow means it is tight. Red means your system will likely struggle.

Step 3 — Download a Q4_K_M variant

In the variant list, click the row labeled Q4_K_M and then click Download. The progress bar at the bottom of the screen fills as the model file arrives. A 7B Q4_K_M model is typically 4.1 to 4.4 GB, so the wait depends entirely on your connection. The file lands in a local models directory that LM Studio manages; you can change this path in Settings if you need it on a different drive.

Step 4 — Load the model

Once the download finishes, click the model name in My Models (second sidebar icon) and then click Load model. A dialog appears with a GPU layer-offload slider. If LM Studio detects a compatible GPU, slide it to the right to offload as many layers as your VRAM allows — this is the single biggest speed improvement available. Click Load. The status bar at the bottom of the window changes from "No model loaded" to the model name with a green indicator when it is ready.

Step 5 — Start chatting

Click the Chat icon (third sidebar item). A session opens with a system prompt field at the top and a message input at the bottom. In the system prompt field, type something concise: You are a helpful assistant. Be direct and accurate. In the message field below, type a question and press Enter or click the send button. The model generates a response token by token. When it finishes, the input clears. You have completed the LM Studio quickstart.

Where to go next

After a successful chat session, three directions are worth exploring: server mode for developer integration, chat presets for saving configurations, and performance tuning for faster inference.

The full tutorial adds a sixth step — enabling server mode and verifying it with a curl command — which matters as soon as you want to connect LM Studio to a code editor, agent framework, or any external tool. The performance page explains how to use the GPU layer-offload slider strategically and how RAM capacity relates to which model sizes are practical. The troubleshooting page covers the five most common problems new users encounter, including what to do when a model fails to load or the GPU goes undetected.

For those evaluating whether local inference fits their use case, AI.gov's public use case catalog provides a useful survey of how public and private organizations are applying AI tools. The NIST AI resources hub covers frameworks for responsible evaluation that apply regardless of which local inference tool you settle on.