Run a Local LLM: One Free App, One 2GB Model

A local LLM is a language model that runs on your own computer instead of a company’s servers, and to run one you install a free app such as Ollama or LM Studio and download a small model. In Ollama, the command ollama run llama3.2 fetches a 2.0GB model and opens a chat in your terminal.

Both tools are free for running models locally, and both say what you type into a local model stays on your machine. Google makes a similar promise for meeting recordings with Foresight, its offline note-taker for the Mac. The catch is that both now offer cloud models too, so it is worth knowing which is which.

Pick a tool

Each maker’s own pages, read 8 October 2026:

	Ollama	LM Studio	llama.cpp
Windows	10 22H2 or newer, Home or Pro	x64 with AVX2, or Arm (Snapdragon X Elite)	Pre-built binaries or a PowerShell install script
Mac	macOS 14 or newer; Intel Macs CPU only	macOS 14 or newer, Apple silicon only	Apple silicon, via Metal
Linux	Install script; x64 and ARM64	AppImage, x64 and ARM64; Ubuntu 20.04 or newer	Install script or build from source
How you use it	App plus terminal commands	Desktop app, with built-in model search	Terminal, plus a web UI when run as a server
Memory guidance	No figure published	16GB+ RAM; on Windows, 4GB+ dedicated VRAM recommended	No figure published
Price for local use	Free, MIT licence	Free plan, $0	Free, MIT licence
Models from	Ollama’s library	Hugging Face, via the app	Hugging Face

For a first try, Ollama and LM Studio do more of the setup for you: each has an installer and its own model catalogue. llama.cpp is an engine others build on, and LM Studio says its runtime has “MLX and llama.cpp under the hood”; llama.cpp’s own quick start is a set of terminal commands.

Run your first model with Ollama

1. Install it. On Windows, run OllamaSetup.exe from ollama.com. Ollama says it installs in your home folder without Administrator rights and needs at least 4GB of disk for itself. On a Mac, open the .dmg and drag Ollama to Applications. On Linux, run curl -fsSL https://ollama.com/install.sh | sh.
2. Start a model. Open a terminal and type ollama run llama3.2. Ollama’s quickstart says it “downloads the model and starts a chat on your computer”. Type /bye to leave.
3. Check where it ran. Type ollama ps. The PROCESSOR column reads 100% GPU, 100% CPU or a split. Our VRAM guide explains what a split means and why it slows answers down.
4. Free the memory. Ollama keeps a model loaded for five minutes after use; ollama stop llama3.2 unloads it at once.

Downloaded models live in ~/.ollama/models on a Mac, /usr/share/ollama/.ollama/models on Linux and C:\Users\%username%\.ollama\models on Windows. Ollama warns that models “can be tens to hundreds of GB in size”; the OLLAMA_MODELS setting moves them to another drive.

Or with LM Studio, without a terminal

1. Install the app from lmstudio.ai. Its home page now leads with Bionic, LM Studio’s agent; the Free plan includes it and local model support.
2. Open the Discover tab (⌘+2 on a Mac, Ctrl+2 on Windows and Linux) and search for a model by name, such as gemma or llama.
3. Pick a download option. LM Studio lists several, named things like Q3_K_S or Q8. Its advice is to “choose a 4-bit option or higher if your machine is capable enough”.
4. Load it and chat. The models folder can be changed under My Models.

Choose a small first model

Download size is the number to read first, because the model has to fit in memory. From Ollama’s library:

Ollama tag	Download	Note
llama3.2:1b	1.3GB	The 1B model
llama3.2	2.0GB	The 3B model; last updated two years ago
gemma4:e2b	4.6GB–7.5GB	The example in Ollama’s quickstart
gemma4	6.6GB–9.5GB	The default tag is the e4b model, not the smallest

The Gemma 4 ranges span two builds: the e2b tags page lists a 4.6GB four-bit build and a 7.5GB MLX build. Beyond these, what fits on the graphics card you already own works through which model sizes suit 8GB, 12GB, 16GB and 24GB cards.

What works offline, and what does not

LM Studio is the more explicit of the two. Its offline page says chatting with models, chatting with documents and running a local server need no internet, and that “nothing you enter into LM Studio when chatting with LLMs leaves your device”. Searching for models, downloading them, fetching new runtimes and app updates do need a connection.

Ollama’s FAQ says “Ollama runs locally. We don’t see your prompts or data when you run locally”, and that it pulls models from the internet. Its pages we read do not use the word offline.

Watch for cloud models. Ollama’s library now carries cloud tags, such as gemma4:cloud, and its FAQ says that for those “we process your prompts and responses”, though it does not store or log them. To rule them out, set OLLAMA_NO_CLOUD=1 and restart Ollama, which also removes its web search. LM Studio’s paid plans, at $20 and $100 a month, are for its cloud-hosted models; the $0 plan runs local ones.

Two things that surprise new users

The conversation is short by default. Ollama’s context-length page gives 4k tokens on machines with under 24GiB of graphics memory, and recommends at least 64,000 for agents and coding tools. Raising it costs memory, which is why local coding assistants often disappoint. Ollama gives the same 64k advice for running the OpenClaw agent on a local model.

The fan gets loud. A model answering keeps the processor or graphics chip busy, and on a laptop that means the fan speeds up. Apple and Microsoft call that rushing-air sound normal under heavy work; what makes a laptop fan loud covers when it is not.

Questions people ask

What is a local LLM?

A large language model that runs on your own computer rather than on a company’s servers. You download the model file once, and a program such as Ollama or LM Studio runs it with your processor and graphics chip. Ollama and LM Studio both say prompts sent to a local model stay on your machine.

How do I run a local LLM on Windows?

Install Ollama with OllamaSetup.exe, which needs Windows 10 22H2 or newer and no Administrator rights, then open a terminal and type ollama run llama3.2. Or install LM Studio, which needs a CPU with AVX2 on x64 PCs, and download a model from its Discover tab.

Can a local LLM run without internet?

Yes, once the model is downloaded. LM Studio says chatting, chatting with documents and its local server need no internet; searching for and downloading models do. Ollama says it runs locally and pulls models from the internet.

Is Ollama free?

Yes. Ollama is MIT-licensed open-source software and running models locally costs nothing. It also offers cloud models, which need you to sign in; Ollama’s FAQ explains how to turn cloud features off.

Ollama or LM Studio for a beginner?

LM Studio is a desktop app with a built-in model search, so nothing is typed into a terminal. Ollama has an app too, but its steps for running a local model are terminal commands. LM Studio does not support Intel Macs; Ollama does, on the CPU only.

What is the smallest model to try first?

In Ollama’s library, Llama 3.2 is 2.0GB and its 1B version is 1.3GB. Start there, then run ollama ps to see whether it loaded onto the graphics chip. Which larger models fit depends on how much memory that chip has.

Where does Ollama store downloaded models?

Ollama’s FAQ gives ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux and C:\Users%username%.ollama\models on Windows. Set the OLLAMA_MODELS environment variable to keep them elsewhere, for example on a larger drive.