LM Studio Tutorial: Getting Started the Right Way

LM Studio turns your computer into a private chat interface for open models. Covers hardware checks, install, your first model, and the local server.

By Galactic News Editorial · Published 2026-11-10T22:28:24+00:00 · Updated 2026-11-10T22:28:24+00:00

LM Studio Tutorial: Getting Started the Right Way

LM Studio turns your own computer into a private chat interface for open language models. You download the app, pick a model from Hugging Face, load it, and chat. No account is needed for local use, and your prompts stay on your machine. This lm studio tutorial walks through the whole setup the right way: check your hardware first, install the app, choose a model that fits, load it with sane settings, and connect your tools through the local server. Follow the steps in order and you will have a working local model in under half an hour.

What LM Studio Is, and What It Is Not

LM Studio is a desktop app for macOS, Windows, and Linux. It downloads open language models, runs them on your hardware, and gives you a familiar chat window. It also serves models through a local API, so your editor and scripts can use them too.

It is not a model itself. Installing the app alone gives you no assistant. You still need to download model weights before you can chat.

Under the hood, LM Studio runs GGUF models through llama.cpp on every platform. On Apple Silicon Macs, it can also run models through MLX, Apple's machine learning framework. The app picks the right engine for your hardware, and you can swap runtimes without reinstalling.

The desktop app is free, including for internal business use. Paid extras exist for cloud inference, but local chat costs nothing and needs no subscription. That makes it a low-risk way to try local models.

LM Studio Tutorial: Check Your Hardware First

Memory decides which models you can run. The official docs recommend 16 GB of RAM. Machines with 8 GB can still run small models, but you should keep the model size and the context length modest.

On macOS, you need Apple Silicon, meaning an M1 through M4 chip, and macOS 13.4 or newer. MLX models need macOS 14 or newer. Intel Macs are not supported at all, so check your chip before you download.

On Windows, you need an x64 CPU with AVX2 support, or an ARM machine such as one with a Snapdragon X Elite chip. The docs recommend at least 16 GB of RAM here too.

On Linux, LM Studio ships as an AppImage and needs Ubuntu 20.04 or newer. Newer Ubuntu releases above version 22 are less well tested, according to the docs, so expect the odd rough edge there.

A GPU with at least 4 GB of VRAM helps a lot. It is not a hard requirement. CPU-only runs work for small models, but they are slower, so keep your first model small if you have no GPU.

Install LM Studio and Run Your First Model

Follow these steps in order. Each one depends on the last.

1. Download the installer from the official downloads page. Pick the build for your operating system. Do not download LM Studio from third-party mirrors.
2. Install it like any normal app. On macOS, drag it into Applications. On Windows, run the installer and accept the defaults. On Linux, mark the AppImage executable with chmod +x and launch it.
3. Open the Discover tab. This is the model browser. It searches Hugging Face directly, so you never leave the app to find a model.
4. Search for a model by name. Try a family you know, such as Llama, Qwen, Gemma, or Phi. Start with a small one, in the 7 to 8 billion parameter range.
5. Pick a quantisation. LM Studio lists variants such as Q4_K_M and Q8_0. The official guidance is blunt: choose a 4-bit option or higher if your machine can run it. Lower bits mean a smaller file and faster replies.
6. Check the size and the memory estimate before you download. The app shows both. A 7 to 8 billion parameter model in 4-bit form is a sensible first pick for most machines.
7. Download the model. You need an internet connection for this step. Large models take a while, so let the download finish before you move on.
8. Open the Chat tab and load the model. Press Cmd + L on macOS or Ctrl + L on Windows and Linux to open the model loader, then select the model you downloaded.
9. Send a test prompt. Ask it to summarise a short paragraph you paste in. If it answers, your setup works.

Tune the Three Settings That Matter

Most problems in LM Studio come from three settings. Get these right and the rest is detail.

Context length. This is how much text the model can hold in memory at once. Recent versions default to 8,000 tokens. Longer context holds more conversation, but it uses more memory. Start at 8,000 and raise it only when you need it.

GPU offload. This controls how much of the model runs on the graphics card. Full offload is fastest. If loading fails or the app stutters, lower the offload until the model fits.

Temperature. This controls how random the output is. Low values, around 0.2 to 0.4, suit factual work and code. Higher values suit creative drafts and brainstorming. Change one setting at a time, so you know what helped.

LM Studio ships presets for common tasks. Start with a preset that matches your job, then adjust a single setting if the output feels off. Random changes to five settings at once teach you nothing.

Serve Models to Your Tools With the Local Server

The chat window is only half the value. The other half is the local server, which lets your tools talk to your model.

Open the Developer tab and start the server. It listens on http://localhost:1234 and speaks the OpenAI API format. Any OpenAI-compatible client can point at http://localhost:1234/v1 and use your local model instead of a cloud one.

The server exposes the endpoints your tools expect, including chat completions, completions, and embeddings. LM Studio also serves an Anthropic-compatible endpoint, so tools such as Claude Code can connect directly with no translation layer.

The lms command-line tool ships with the app. Run lms server start to start the server from a terminal, and lms load to load a model with flags for context length and GPU offload. You need to run LM Studio at least once before lms works.

For machines with no display, such as a Linux GPU box or a build server, use llmster. It is the same inference engine packaged as a headless daemon, with no window at all.

Chat With Your Documents, Offline

You can drag a document into a chat and ask questions about it. The app processes the file on your machine, so nothing you upload leaves the application. This works fully offline once the model is downloaded.

Keep documents small at first. Large files eat context fast, and a long document can push useful conversation out of the window. Split big files into sections and ask about one section at a time.

Remember that the model does not learn from your chats. It only knows what is in the current conversation, plus any documents you attach. Close the chat and that context is gone.

Mistakes to Avoid

Downloading a model bigger than your RAM. The app shows a memory estimate for each variant. Trust it. If the estimate exceeds your free memory, pick a smaller model or a lower quantisation.

Skipping the quantisation choice. A 4-bit model often runs far faster than an 8-bit one, with a small quality cost. For a first run, 4-bit is usually the right call.

Expecting cloud-level speed on a laptop CPU. Local models are slower than hosted ones. Small models answer fast. Large ones crawl. Match the model to your hardware, not to your ambitions.

Treating downloaded weights as licence-free. Each model has its own licence, separate from the app. Check it before you use a model for commercial work.

Using a cloud feature and expecting offline privacy. Local chat and local document processing stay on your machine. Cloud models and web search change that, so check which mode you are in before you paste anything sensitive.

Bottom Line

LM Studio is the easiest path to a private local model. Check your hardware, install the app, pick a model that fits your memory, load it, and chat. Turn on the local server when your tools need it, and keep your first model small. If the first model feels slow, the fix is almost always a smaller model or a lower quantisation, not a new app.