Skip to content

THE FIELD JOURNAL

How to Run Llama Locally on Your Computer

By Local AI Scout · Beginner guide

Paper-cut illustration of a small llama standing on top of a desktop computer tower next to a laptop, a desk lamp and a notebook

By Local AI Scout · Local · Models · Guide · Docs-only · Checked November 18, 2026

Paper-cut illustration of a small llama standing on top of a desktop computer tower next to a laptop, a desk lamp and a notebook

Short answer: Yes, you can run a Llama model on your own computer if you pick a size that actually fits. "Llama" is a family, not a single download. On Ollama's library pages we opened November 18, 2026, the default llama3.2 tag (a 3B model) is listed at 2.0GB, llama3.2:1b at 1.3GB, llama3.1:8b at 4.9GB, llama3.1:70b at 43GB, and Llama 4 Scout (llama4 / llama4:16x17b) at 67GB. Start with a small tag, then use our memory formula before any large download. Llama has its own license with conditions; our commercial-use guide walks through those in plain English. We did not download or run these models.

This guide follows one reader: someone who was told to "just run Llama" and doesn't know whether that means a 2GB file or a 67GB one.

Llama is a family, not one file

Meta's Llama page, opened November 18, 2026, groups several lines: Llama 4 (natively multimodal mixture-of-experts models), the Llama 3 family (including 3.1, 3.2 and 3.3), and older Llama 2 models. The useful question for a home computer is not "can I run Llama?" It is "which named file am I about to download, and how big is it?"

A short way to read the names:

  • A figure like 3B or 8B is the model's size in billions of parameters. Smaller usually means a smaller file and a better chance of fitting a laptop.
  • Llama 4 Scout and Llama 4 Maverick are mixture-of-experts models. Ollama's Llama 4 readme lists Scout as 109B parameters with 17B active, and Maverick as 400B with 17B active. The file still has to be stored, even if only some of the experts run on each token. See our models hub for that distinction.
  • Tags like Q4_K_M are compressed "quantized" copies of the same model. Our GGUF guide explains the labels.

Our models page, as of its October 26, 2026 check of Ollama's library, listed Llama 3.1 first by pulls. That is a snapshot from that page, not a ranking we re-measured today.

Sizes we opened on November 18, 2026

Figures below are the download sizes shown on Ollama's library pages that day. They are not speed tests, and they are not a promise the model will feel fast on your machine. Always re-check the library page before you download.

Ollama commandWhat the library page called itSize shownContext shown
ollama run llama3.2 or ollama run llama3.2:3bLlama 3.2, 3B (default)2.0GB128K
ollama run llama3.2:1bLlama 3.2, 1B1.3GB128K
ollama run llama3.1 or ollama run llama3.1:8bLlama 3.1, 8B (default / latest)4.9GB128K
ollama run llama3.1:70bLlama 3.1, 70B43GB128K
ollama run llama3.1:405bLlama 3.1, 405B243GB128K
ollama run llama4 or ollama run llama4:scoutLlama 4 Scout (llama4:16x17b)67GB10M listed on the tag
ollama run llama4:maverick or llama4:128x17bLlama 4 Maverick245GB1M listed on the tag

The 70B, 405B and Llama 4 files are in "desktop PC with a large disk and plenty of memory" territory, not "try this on the family laptop after dinner." A 2.0GB or 4.9GB file is the realistic first attempt. Even then, add conversation memory and a little room for the app and the operating system — that is what the memory formula is for. We are not claiming a 4.9GB file will fit comfortably in 8GB of RAM.

A first run with Ollama

This is the path we can document from pages we opened. We have not run the commands.

  1. Install Ollama if you don't already have it. Our Ollama install guide covers Windows, Mac and Linux and the first errors.
  2. In a terminal, type one of the small-tag commands from the table, for example ollama run llama3.2. That command is the one on Ollama's Llama 3.2 library page.
  3. Wait for the download to finish, then try a short prompt. If the machine starts swapping heavily or the app runs out of memory, quit and try a smaller tag, or check the formula page before you try again.
  4. Remove a model you don't need with the steps in our guide to deleting Ollama models — useful after a 67GB curiosity download.

Treat any default context window as a setting, not as something you have to fill. A 128K window is large; it costs memory. Our context-length guide explains why that matters.

Other apps, and the license

LM Studio and llama.cpp can also run Llama files in GGUF form. We did not re-check LM Studio's catalog labels today ‡, so start with Ollama's library names if you want a command you can compare with this page.

Downloads should come from a place you trust. Our model-download safety guide covers GGUF and other formats.

Open weights are not the same as "free for any business use." Our commercial-use guide summarizes the Llama license conditions we opened there (including the "Built with Llama" display rule and Meta's permission threshold for very large services). Read the license that ships with the exact tag you download. This is general information, not legal advice.

Choose a small Llama tag if…

  • You want to try the family on a laptop, and you will start with Llama 3.2 (2.0GB on the page we opened) or Llama 3.1 8B (4.9GB) and check memory first.
  • You are happy to stay in a chat app such as Ollama, without chasing the newest 67GB name.

Choose a different plan if…

  • You heard "run Llama 4" and only have a 16GB laptop. The Scout tag we opened is a 67GB download.
  • You need a clearly permissive license such as Apache 2.0 for a client project. Check the commercial-use guide and the model card before you commit.
  • You already have a first model you like in our Ollama picks. You don't have to run Llama just because the name is familiar.

Frequently asked questions

Can I run Llama 4 on a laptop?

Only if the laptop can hold the file and the working memory. Ollama's library page, opened November 18, 2026, listed the default Llama 4 / Scout tag at 67GB and Maverick at 245GB. That is much larger than Llama 3.2 at 2.0GB or Llama 3.1 8B at 4.9GB. We did not test Llama 4 on a laptop.

What is the smallest Llama I can try first?

On the Ollama pages we opened, Llama 3.2 1B is listed at 1.3GB and the default Llama 3.2 3B tag at 2.0GB. Those are the smallest official Llama tags we saw in that library. Check today's listing and the memory formula before you download.

Do I have to use Ollama?

No. Ollama is the catalog we opened for file sizes and commands. LM Studio and llama.cpp can also run Llama files. We did not re-check other apps' exact catalog names today.

Is Llama free to use in a business?

Not automatically. Meta publishes separate Llama licenses. Our commercial-use guide summarizes conditions we opened on those licenses, including display and permission rules that can apply if you distribute a product that contains Llama. Read the license for the tag you download. This is not legal advice.

Why do some pages still talk about Llama 3.1 if Llama 4 exists?

They are different lines. Meta's page lists Llama 4 and the Llama 3 family side by side. Llama 3.1 8B is still the smaller everyday tag in Ollama's library; Llama 4's listed downloads are much larger. "Newest" and "fits my computer" are not the same test.

Sources

  • Meta, "Get started with Llama" (https://dev.meta.ai/llama), opened November 18, 2026.
  • Ollama library, Llama 3.2 (https://ollama.com/library/llama3.2), opened November 18, 2026: default 3B tag 2.0GB; 1B tag 1.3GB; 128K context.
  • Ollama library, Llama 3.1 (https://ollama.com/library/llama3.1), opened November 18, 2026: 8B / latest 4.9GB; 70B 43GB; 405B 243GB; 128K context.
  • Ollama library, Llama 4 (https://ollama.com/library/llama4), opened November 18, 2026: llama4 / llama4:16x17b (Scout) 67GB; llama4:128x17b (Maverick) 245GB. Readme: Scout 109B MoE with 17B active; Maverick 400B MoE with 17B active.
  • Our live guides, checked November 18, 2026: models hub, how much memory a local AI model needs, can you use local AI models commercially?.

How we checked: a documentation-based guide, not a hands-on test. We did not download or run these models. File sizes are as shown on the Ollama pages on November 18, 2026 and can change. ‡ marks something we could not confirm on an official page. See our editorial policy.

Keep exploring: Apps · Models · All guides

THE SCOUT LETTER

Good finds.
Less noise.

Useful local AI apps, practical guides and the updates worth your time. A thoughtful note in your inbox.

Free to join. Unsubscribe with one click.