With a Laptop: 10 LM Studio Steps to Run Free AI

Free at home and at work

Your laptop runs zero-dollar AI that works offline: ten LM Studio steps, simplest first. Here's how you start. Number ten: LM Studio is a free download that runs on your own computer. You can use it both at home and at work.

You type a question, and the answer comes from your own machine. Your messages and chats stay saved on your computer. Once a model is downloaded, it keeps working with no internet at all. It handles three jobs: chats, your own documents, and a server for other apps.

The app itself comes in English and more than thirty other languages. By our count, that's at least thirty-three languages for the app. It also finds and downloads models for you. It's the easiest way in if you're new to running AI yourself.

The link is in the description. Up next, installing it.

Install LM Studio

Number nine: the download page has an installer for Mac, Windows and Linux. The site offers two apps, and this video uses the one named LM Studio. You install the latest version, then check that your computer can run it. In User mode, it shows only the chat and sets up everything else for you.

That's the best choice for beginners, or anyone happy with the default settings. On a Mac, it needs macOS fourteen or newer, on one of the newer M-series chips. Sixteen gigabytes of memory is recommended, and eight may still work with smaller models. By our count, the recommended sixteen gigabytes is twice the eight that may still work.

If step-by-step guides like this help you, subscribe. Every official page is linked below. Coming up, picking a model.

Pick a model that fits your memory

Number eight: the Discover tab lists models, picked for you or found by search. Command or Control plus two opens that tab from anywhere. Mistral's Ministral models come in three sizes: three, eight and 14 billion parameters. The smallest needs at least two gigabytes of memory, and the largest up to ten.

By our count, the largest needs up to five times the memory of the smallest. Their open license allows commercial and non-commercial use. Mistral's 7 billion model needs at least four gigabytes of memory. DeepSeek's reasoning models start at four gigabytes, and the largest needs up to forty.

With sixteen gigabytes of memory, DeepSeek's smaller seven- and 8 billion versions run. The app warns you before loading a model that may not fit. Then, downloading it.

Download the model

Number seven: Mistral's 7 billion model is a 4.1 gigabyte download. The smallest Ministral is two gigabytes, and DeepSeek's small one is 4.3. By our count, all three downloads come to 10.4 gigabytes. Each model offers several versions, with names that start with a Q.

They're copies of the same model, at different levels of fidelity. The Q stands for quantization, which shrinks the file and gives up some quality. A four-bit version or higher is the pick, if your computer can run it. Downloading needs a steady, fairly fast internet connection.

Models from outside the app can be brought in, too. Under My Models, the folder where models are kept can be changed. Now, your first chat.

Your first chat

Number six: in the Chat tab, the model loader opens a model you downloaded. Loading it sets aside part of your computer's memory for the model. Then you talk with it back and forth, like any chat window. By our count, Mistral's 4.1 gigabytes leaves at most 11.9 of sixteen gigabytes.

The plus button, or Command or Control and N, starts a new chat. Chats can go in folders, and two can sit side by side. The model doesn't learn from your chats; it only knows what's in them and its settings. Chats are saved as files on your computer, and you can search them all.

They also export to a PDF, markdown or plain text. Next, your own documents.

Ask about your own documents

Number five: you can attach your own files to a chat and ask about them. It reads three kinds of files: word-processor files, PDFs and plain text. That gives the model extra context for what you ask. Context is the model's working memory, and it has a maximum size.

It's measured in tokens, and one token is often about three quarters of a word. If a document is short enough to fit, the whole file goes into the conversation. If it's very long, the app pulls out the most relevant bits for the model instead. That's called retrieval-augmented generation, or RAG for short.

It can do the same across several documents at once. Say Maya attaches a three-thousand-word report and asks what it covers. By our count, that report is about 4,000 tokens. Sometimes it works really well, and sometimes it needs some tuning and experimenting.

The tip is to put as much context as you can into your question. Mentioning the terms and ideas you expect in the document helps it find the right part. A document you drag in stays on your machine, and all the processing happens offline. The Ministral models can hold up to 256,000 tokens.

By our count, that's about 192,000 words at the most. The context length you pick before loading also changes how much memory it needs. Up next, setting its behavior.

Set its behavior with a system prompt

Number four: a system prompt sets the model's behavior, or the instructions it follows. The docs' own example is short: always answer in rhymes. A large System Prompt editor opens with Command or Control and E. Presets bundle a system prompt and other settings into one setup you reuse across chats.

That makes Presets your own prompt library. Every setting in the Advanced Configuration sidebar can be saved in a named preset. A preset can always use a certain temperature, top P or max tokens. By our count, one preset can hold at least four things: the prompt and those three settings.

You switch between uses like creative writing, brainstorming or reasoning. Temperature is the randomness in which words the model picks. Zero is deterministic, and higher values increase creativity, on a scale up to one. Max tokens caps how much the model writes in one answer.

Say Maya saves a preset that writes product descriptions in her shop's voice. Next time, her prompt and settings are ready in that same preset. Presets are small files you can share, or import from others by a link. If an option is missing, Developer mode, in Settings, then Developer, shows them all.

Developer mode also includes keyboard shortcuts and development features. Each model can also keep its own default load settings, set from My Models. In most cases, the prompt template never needs changing. Then, keeping it private and offline.

Private and offline

Number three: nothing you type into a chat with a downloaded model leaves your device. Once a model is on your machine, it runs locally, and you're good to go offline. In general, the app doesn't need the internet to work. That includes chatting, chatting with documents and running a local server.

Five things need it: searching for models, downloading them, and the catalog's download options. Downloading runtimes, the engines that run models, and checking for app updates make five. Say Maya drafts a client email on a train with no signal, and the model still answers. By our count, that's three things that work offline, and five that need internet.

None of the five is about your chats; they're about getting models, engines and updates. The app ships with a starting copy of the model catalog built in. Opening the app checks for updates over the network. That update check collects no personal identifiers and doesn't track you.

Without a connection, you just can't update through the app's own updater. Its privacy policy also says the app includes no telemetry. Beyond paid cloud features, the makers hear from the app only on model searches, downloads and update checks. By our count, that's two occasions, both about getting files, and neither is about what you type.

Search queries for models are anonymized, and they're not tied to you. Your information is never sold to third parties. Coming up, using it from other apps.

Use it from other apps

Number two: a local server lets your other apps use the models on your computer. In the Developer tab, the Start server switch turns it on. A terminal command can start it too. Once it's running, it listens at a local address, on port twelve thirty-four.

Your own apps and scripts can then use your local models through its REST API. Ready-made libraries for Python and TypeScript connect to it too. The docs list six endpoints for it, from listing models to chat completions. By our count, that's one switch, one port and six endpoints.

By default, the server asks for no key. API tokens can be switched on, so only requests with a valid one are accepted. That switch sits in the Developer page, under Server Settings. Each API token gets a name and its own permissions.

A new token is shown only once, so it gets copied right away. Your other apps can even use a model without you loading it first. It works on your own computer, or across your local network, with no internet. One option lets other devices on your network use it, too.

Say Maya runs a small script that sends her product notes to the local model. Each request goes to her own computer at port twelve thirty-four, not to a faraway server. If the app isn't running, a terminal command starts it automatically. Last, keeping it running well.

Keep it running well

Number one: on Mac and Windows, a built-in updater keeps the app current. When a new version is out, a notification lets you update now or later. The Linux updater is still in the works. Runtimes, the engines that run the models, update without a full app update.

Command or Control, Shift and R opens them. When another app loads a model on its own, an idle timer can unload it later. The timer resets with every request, so the model won't disappear while you use it. By default, that idle time is sixty minutes.

The docs' example sets it to 300 seconds. By our count, 300 seconds is five minutes, and the default is twelve times longer. Auto-Evict, on by default, keeps at most one of those models loaded at a time. Models you load from a terminal stay loaded until you unload them.

Any model can be unloaded from a terminal too, one or all at once. A terminal dry run prints how much memory a model would need, without loading it. If the estimate looks too cautious, you can still try loading the model anyway. Its online community answers questions, and bugs go to its bug tracker.

Say Maya's apps call three different models of 4.1 gigabytes each. By our count, all three loaded at once would take 12.3 gigabytes. With Auto-Evict on, only one stays loaded, 8.2 gigabytes less. Now that your AI runs free, which sites pay you to train AI?

Click the video on screen now.