How to Run AI on Your Computer Without Internet

Most AI assistants live on someone else’s servers. Every question you type travels over the internet, and the answer travels back. That is fine for many tasks, but not for all: you might be on a plane, in a place with weak internet, or working with notes and documents you would rather not upload anywhere.

The alternative is a local model, an AI that runs on your own computer. One of the easiest ways to try it is LM Studio, a desktop app that downloads open models and lets you chat with them. This guide shows how to set it up, what “offline” really means, what hardware you need, and what to expect from the results.

One honest note up front: you need internet once, to download the app and at least one model. After that, chatting can work without a connection. Everything below comes from LM Studio’s own documentation and site where it loaded, and I mark what I could not confirm. I did not run these steps myself for this article.

What LM Studio Is

LM Studio’s documentation describes it as an app for downloading and running local models, with a chat interface, support for MCP servers, search and download of models from Hugging Face, and the ability to serve models on OpenAI-style endpoints on your machine. It supports macOS, Windows, and Linux. The company’s home page now also promotes an agent called Bionic, built on the LM Studio runtime with MLX and llama.cpp. You do not need Bionic to follow this guide.

The practical point: instead of sending your text to a company’s server, LM Studio loads a model into your computer’s memory and runs it there.

What You Need: Hardware and System Requirements

LM Studio’s system requirements page lists:

The page does not list disk space requirements. Models are large files, so leave plenty of free storage. Check how much room a model needs before you download it.

A rough rule of thumb (mine, not LM Studio’s): a model’s file size is roughly its number of parameters times the bits per weight, divided by eight. A 7-billion-parameter model at 4-bit is around 3.5 GB before overhead. The model has to fit in memory with room to spare, which is why more RAM or GPU memory opens up larger and better models.

What Works Offline, and What Doesn’t

LM Studio’s offline operation page is clear about the split.

Works without internet:

• Chatting with models you have already downloaded. The page says chat input stays on your device.
• Chatting with documents. Documents you add are processed locally and do not leave the app.
• Running the local server, which serves OpenAI-style endpoints on localhost or your local network without internet access.

Needs internet:

• Searching for models in the Discover tab, which makes network requests.
• Downloading new models, which needs a stable, reasonably fast connection. You can sideload model files obtained elsewhere.
• Downloading runtimes such as llama.cpp engines.
• Checking for app updates through the in-app updater.

So the sequence is: get everything while you are online, then switch your connection off.

Step-by-Step: Set Up LM Studio and Chat Offline

LM Studio’s getting started page gives a five-step outline. Here is that outline expanded.

Step 1: Check your computer

Compare your machine with the requirements table above. If you have a Mac, confirm it is Apple Silicon. If you have a Windows PC, confirm it has the RAM you need.

Step 2: Install LM Studio

Download the installer from the LM Studio download page. When I read the page, it showed a direct Windows (x64) download, and its metadata mentioned Mac and Linux but I did not see installer links for them in the text I could read. If your platform’s link is not obvious, check the page in your browser. Run the installer and open the app.

Step 3: Open the Discover tab and find a model

In the Discover tab you can pick a curated option or search by name. The download documentation gives examples such as llama, gemma, or a specific user/model string, and says you can paste a full Hugging Face URL into the search bar.

If you are new, start small. A smaller model will download faster, load faster, and run on more hardware. You can move up later.

Step 4: Pick a quantization

Many models have several versions with names like Q3_K_S or Q_8. LM Studio explains that these are copies of the same model at different fidelity levels. The “Q” stands for quantization, which shrinks the file but reduces some quality. The documentation advises choosing a 4-bit option or higher if your machine can handle it.

In plain terms: a lower number means a smaller, faster, less accurate model, and a higher number means a bigger, slower, more accurate one.

Step 5: Download and load the model

Download your choice while you are online. Then, in the Chat tab, open the model loader, select the downloaded model, and optionally adjust load settings. The documentation notes that loading allocates memory for the model’s weights in your computer’s RAM.

Step 6: Chat

Once the model is loaded, type a message. The first reply may take a moment.

Step 7: Test it offline

Turn off Wi-Fi or enable airplane mode, then send a message. If it answers, your setup works. Do this before a trip, not at the airport.

Step 8: Keep a copy of what you need

Because model search and downloads require internet, download any model you might want while you are connected. Updates to the app also require a connection, so plan to update when you are back online.

Chat With Your Own Documents

LM Studio lets you attach .docx, .pdf, and .txt files to a chat so the model has extra context. How it works depends on length. A short document that fits in the model’s context window is added to the conversation in full. A very long document, or several, is handled with retrieval-augmented generation (RAG), which pulls out the passages most relevant to your question and gives those to the model. The page says RAG sometimes works well and sometimes needs tuning. It suggests including as many relevant terms and ideas as you can in your question.

The document page itself does not state limits on the number or size of files. The separate offline page says documents are processed locally. For a deeper look at the risks of uploading files to cloud tools instead, see our checklist in Free AI Tools for Translating and Summarizing Documents.

Using LM Studio as a Local Server

If you code or use apps that can talk to an AI service, LM Studio can act as a local one. The developer documentation shows a command-line example, lms server start --port 1234, with requests sent to http://localhost:1234. The examples include an authorization header with an API token variable. I did not find a stated default port, so use the port you set. This is optional and more technical than chatting.

Privacy: What Stays on Your Device

LM Studio’s privacy policy says that with local use your data stays on your device: none of your messages, chat histories, and documents are transmitted from your system, the application includes no telemetry or user-specific tracking, and prompts and responses are not retained.

It also lists what can leave your computer:

• Update checks send the app version, build, and operating system, and your IP address is visible to the content delivery provider.
• Model search and downloads send anonymized search queries to the company.
• Cloud services, only if you buy and use features such as cloud models or web search, where prompts and files are processed transiently by providers under zero data retention or similar terms.

That is the company’s own description. If privacy is your main reason for going local, keep the cloud features off and use the app offline once your models are downloaded. For comparison, see how a cloud assistant handles private chats in our look at ChatGPT’s Locked Chats.

What It Costs

LM Studio’s pricing page lists a $0 Free plan that includes running local models on your machine. Paid plans cover cloud-hosted models and higher usage: Bionic+ at $20 a month and Pro at $100 a month. You do not need a paid plan to run models offline.

The pricing page does not say whether personal or work use is allowed on the Free plan. If you plan to use LM Studio for a business, read the site’s terms of use first.

What to Expect From Local AI

Be realistic about the trade-offs:

• Speed depends on your hardware. An older laptop with little memory will respond slowly or cannot load larger models.
• Quality depends on the model. Smaller local models are generally less capable than the largest cloud models, though that gap varies by task and model, so test before you rely on one.
• No live information. An offline model knows only what it learned in training and what you give it. It cannot look things up.
• Mistakes still happen. Local models can state wrong facts confidently. Check anything important.
• Disk space adds up. Each model you try takes room, so delete the ones you do not use.

Common Mistakes to Avoid

Expecting to download a model offline. Downloads need internet, so get what you need first.

Choosing a model that is too big. If the app is slow or fails to load, pick a smaller one or a lower quantization.

Going too low on quantization. LM Studio advises 4-bit or higher where your machine can run it.

Assuming offline means perfectly private. Update checks and model searches still contact the company if you use them.

Forgetting to test offline. Disconnect and try a chat before you travel.

Trusting document answers blindly. RAG can miss relevant passages. Verify quotes.

Frequently Asked Questions

Can I really use AI without internet? Yes, after you have installed LM Studio and downloaded a model. Chatting and chatting with documents run locally.

Is LM Studio free? Its pricing page lists a $0 plan that includes local models. Paid plans are for cloud models and extra usage.

How much RAM do I need? LM Studio recommends 16GB or more on Mac and Windows. An 8GB Mac may work with smaller models.

Does it work on an Intel Mac? No. The requirements page says Apple Silicon is required.

Does LM Studio send my chats anywhere? Its privacy policy says messages, chat histories, and documents are not transmitted from your system when you use it locally.

Can it read my PDFs? Yes. It supports .docx, .pdf, and .txt files, using RAG for long documents.

The Bottom Line

Running AI offline is no longer a project only for developers. With LM Studio, the steps are simple: check your hardware, install the app, download a model while you are online, load it, and disconnect. You get a private, subscription-free assistant that works on a plane or in a basement, at the price of slower speed and, usually, a less capable model than the biggest cloud services. Start with a small model, test it offline, and move up only if you need more. If you want ideas for what to do with an assistant like this, our roundup of free AI note-taking and study apps is a good starting point.

Sources

• LM Studio
• LM Studio: system requirements
• LM Studio: offline operation
• LM Studio: get started
• LM Studio: download an LLM
• LM Studio: chat with documents
• LM Studio: privacy policy
• LM Studio: pricing
• LM Studio: download page
• LM Studio: developer documentation