Blog / Build with AI /

LM Studio explained for beginners

LM Studio is a desktop app that lets you download an AI language model and chat with it on your own computer. No account, no subscription and, once the model is downloaded, no internet connection needed. It’s the nearest thing to running your own private ChatGPT-style assistant.

If you’ve read about running an LLM on your own computer and wanted a friendlier route than the command line, this is the kind of tool that gets you there.

The analogy

Think of a cloud chatbot as a restaurant. You send your order, the kitchen (someone else’s powerful servers) cooks it, and the food comes back. It’s convenient, and the kitchen sees every order.

LM Studio is a home kitchen. You buy the ingredients once (the model files), cook on your own hob (your computer’s chip and memory), and nobody else sees what you’re cooking. It’s slower, the recipes are smaller, and you do the washing up. But it’s yours.

What the app does

LM Studio is made of a few parts:

  • A model browser. You search for models, view their size and download them. Many come from the open-source community, such as Meta’s Llama family, Mistral’s models and Google’s Gemma.
  • A chat window. It looks much like any chatbot. Pick a model, type, read the answer.
  • A local server. You can switch on a server that behaves like the OpenAI API, but runs on your machine. Other tools and scripts you write can talk to it without any code changes beyond the address.
  • Settings. Controls for things like how creative the model is (temperature), how much text it can read at once (context length) and how much work goes to your graphics card.

It runs on Windows, Mac and Linux. On Macs with Apple silicon, it runs well, thanks to the unified memory design.

What you need

The thing that matters most is memory. A model has to fit into your computer’s RAM (or your graphics card’s memory) to run at a useful speed.

Rough guidance:

  • Small models, the kind with a few billion parameters, can run on many ordinary modern laptops.
  • Medium models need a machine with a good amount of RAM, and run slower without a decent GPU or Apple silicon chip.
  • Large models are for powerful workstations.

You’ll also need disk space. A single model file can be several gigabytes, and you’ll probably end up with a few.

Don’t memorise specs. The app shows the file size of each model, and many listings indicate whether it’s likely to run on your hardware. Start with a small one.

Understanding model names

Names look like code, so a quick decoder helps.

  • The size (7B, 8B, 70B). Billions of parameters. Bigger models are usually more capable and need more memory.
  • Quantisation (Q4, Q5, Q8). A way of shrinking a model so it fits on ordinary computers. Lower numbers mean smaller files and slightly lower quality. Q4 or Q5 versions are a popular balance.
  • Instruct or chat. These versions are trained to follow instructions. For a normal chat, pick these, not the “base” versions.
  • GGUF. The file format LM Studio uses.

Your first ten minutes

  1. Download and install LM Studio from its official site.
  2. Open the search tab and look for a small, well-known instruct model (a 7B to 8B Llama or Mistral model in a Q4 or Q5 size is a common starting point).
  3. Download it. Be patient; it’s a big file.
  4. Open the chat tab, choose the model and wait for it to load.
  5. Ask something simple, then something harder.

Try tasks where you can check the result: summarise this paragraph, rewrite this email politely, explain what this error means, suggest names for a tool.

Why use it

  • Privacy. Your text stays on your machine. That matters for client documents, internal notes or anything you wouldn’t paste into a website.
  • No per-use cost. After the download, you can run as many prompts as you like.
  • Offline work. Useful when your connection is bad.
  • Experimentation. You can try different models and settings, which helps you understand how they behave.
  • A free test bench for builders. The local server lets you develop and test an app without spending on API calls.

Why you might not

Be honest about the trade-offs.

  • Quality. The models you can run on a laptop are generally weaker than the best cloud models, especially at reasoning, long documents and coding.
  • Speed. Answers can be slow without suitable hardware.
  • Battery and heat. Running a model works your computer hard.
  • Setup and upkeep. You choose models, manage downloads and learn the settings.
  • Hallucinations. Local models make things up, just as cloud ones do. Smaller ones often more so.
  • Licences. Models have licences. Some are open for commercial use and some are not. Read the model card before using one for business work.

Using the local server

For builders, this is the best feature. In LM Studio you can start a local server, which exposes an address on your computer. Code written for the OpenAI API can be pointed at it instead, so a script you wrote for the cloud runs locally.

A tiny example, assuming the server is running on its default port:

from openai import OpenAI

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

reply = client.chat.completions.create(
    model="local-model",
    messages=[{"role": "user", "content": "Give me three names for a recipe scaler tool."}],
)
print(reply.choices[0].message.content)

The details depend on your version and setup, so treat this as a shape and check the app’s own instructions. The point is that you can test a tool on your own machine at no cost per request.

Keep that server private. By default it should only accept connections from your own computer. Don’t expose it to the internet without understanding what you’re opening up.

Where it fits for someone running websites

Practical uses include drafting alt text suggestions for a batch of images, tidying product descriptions, classifying support emails or testing prompts before you spend money on an API. Ollama explained for beginners covers a similar tool with a command-line focus, if you prefer that style.

For tasks where the model must be accurate, such as legal text, medical information or anything customer-facing, do not rely on a small local model without checking.

Common myths

  • “Local means perfectly private.” The app runs offline, but downloads models from the internet, and you should still read what any app collects.
  • “Local models are as good as the best cloud models.” Not on typical hardware.
  • “You need a gaming PC.” Small models run on ordinary modern laptops. You only need more power for bigger ones.

Key takeaways

  • LM Studio runs AI models on your own computer, with a friendly app.
  • Memory decides which models you can run; start small.
  • Model names tell you size, quantisation and whether it’s tuned for chat.
  • The local server is a free place to test tools.
  • Expect weaker answers than top cloud models, and check what it says.

What’s the first document you’d want an AI to read without it ever leaving your laptop?