Aiodic

Back to the table

60 Model family La Llama Model Families

Llama

Meta's openly available models.

What it is

Llama is a family of AI models made by Meta, the company behind Facebook, Instagram, and WhatsApp. Like GPT, Claude, and Gemini, it is a language model: software trained on enormous amounts of text so that it can answer questions, write, summarise, translate, and help with code.

What made Llama different is how Meta released it. OpenAI, Anthropic, and Google mostly keep their best models on their own computers and let you use them through an app or an API, a way for software to talk to other software. Meta, from 2023 to 2025, let people download the Llama model files themselves. Those files contain the model's weights, the billions of learned numbers that make it work, which are explained in Parameters. Once you have the weights, you can run the model on your own computer or server, change it with fine-tuning, and build it into your own products, within the terms of Meta's licence.

That is why the name keeps coming up across this table. Llama was the family that made "run your own AI" a normal idea for developers, hobbyists, and businesses.

The generations at a glance

The dates below come from Meta's own announcements. Model names and sizes are given in billions of parameters, written as "B".

In March 2025 Meta said Llama models had been downloaded more than one billion times. That is Meta's own count, and one person can download a model many times, but it gives a sense of scale.

What "open" means here, and what it doesn't

Meta has long called Llama "open source". Many people who work on open-source software disagree, and the difference matters if you plan to use Llama for real work.

Open-source software traditionally means anyone may use, study, change, and share it, for any purpose, with no one excluded. Llama comes with the weights, but under Meta's own Llama Community License Agreement, which adds conditions. At the time of writing, the Llama 4 licence says, among other things:

To download Llama from Hugging Face, the main website where AI models are published, you must also accept the licence and share your contact details with Meta.

In February 2025 the Open Source Initiative, which maintains the long-standing Open Source Definition, said Llama's licence still doesn't qualify because it restricts who may use the model and for what. The more precise term many people now use is open weights: you get the model files, under a licence with conditions, and without the training data. For comparison, OLMo publishes its training data and code too.

Why it matters

It made AI something you can own a copy of. With a chatbot service, the company can change the model, raise the price, or switch it off, and your data travels to its computers. With downloaded weights, the model on your machine stays the same until you replace it, keeps working without an internet connection, and never sends your text anywhere. For businesses handling client records, and for anyone who cares about privacy, that is a real difference.

It started a whole ecosystem. Because anyone could build on Llama, many variants appeared on Hugging Face: models fine-tuned for medicine, law, coding, or particular languages, and smaller compressed versions that run on modest hardware. Free tools such as Ollama and LM Studio grew up largely around running Llama and its successors at home. Other companies followed with their own open-weight families, including Mistral, DeepSeek, Qwen, and Google's Gemma.

Billions of people have met it without knowing the name. For most people, Llama wasn't a download at all. From April 2024 it powered Meta AI, the assistant inside WhatsApp, Instagram, Facebook, and Messenger, found in the search bar and in chats. In May 2025 Mark Zuckerberg said Meta AI had reached about one billion monthly users across Meta's apps. As with every company's user figures, Meta chooses how to count them.

It fuels the debate over open AI. Supporters say open weights spread the benefits of AI, let researchers study models independently, and stop a few companies controlling the technology. Critics say that once weights are public, safety rules can't be enforced and anyone can strip out a model's guardrails. Llama has been the test case for both arguments, as the examples below show.

It shows that openness is a business decision, and can change. Meta opened Llama because it suited Meta's strategy at the time. When that strategy changed in 2025 and 2026, so did the plans. Anyone building on an open model should expect that the next version may come with different terms, or not come at all.

Real-world examples

Europe's missing vision models (2024 to 2025)

In June 2024, Meta paused its plan to train AI on public posts from Facebook and Instagram users in the European Union, after its lead EU privacy regulator, Ireland's Data Protection Commission, asked it to. A month later, on 17 July 2024, Axios reported that Meta would hold back its coming multimodal models (ones that handle images as well as text) from the EU. Meta's statement said it would release a multimodal Llama model "over the coming months, but not in the EU due to the unpredictable nature of the European regulatory environment".

When Llama 3.2 arrived on 25 September 2024, the restriction was written into its acceptable use policy. The rights to the multimodal models were "not being granted" to individuals living in the EU or to companies with their main place of business there. The small text-only models were not affected. And there was an important exception: the restriction "does not apply to end users" of a product that uses those models. So a French café owner could use an app built by an American company on Llama 3.2 Vision, but a French startup could not legally download the model and build that app itself. The Llama 4 policy repeats the same clause.

Meta's assistant followed a similar path. Meta AI, running in the US since 2023, only began rolling out across 41 European countries in March 2025, starting with a text chat. The Open Source Initiative cited the EU exclusion as one reason Llama's licence isn't open source. The lesson for a reader anywhere: "free to download" and "free for you to use" are separate questions, and the answer can depend on where you live.

ChatBIT and the limits of a usage policy (November 2024)

On 1 November 2024, Reuters reported on a June 2024 paper by six Chinese researchers, including two from institutions under the People's Liberation Army's Academy of Military Science. They described ChatBIT, built by fine-tuning a 13-billion-parameter Llama model on military dialogue records, and said it was optimised for question-answering in the military field. Llama's acceptable use policy bans military uses. Meta's director of public policy, Molly Montgomery, said any use of its models by the People's Liberation Army was "unauthorized and contrary to our acceptable use policy".

That statement was accurate, and it was also all Meta could do. Once weights are downloaded, the company has no switch to turn a copy off, so the policy is a legal promise, not a technical lock. Three days later, on 4 November 2024, Meta's president of global affairs, Nick Clegg, announced that Meta would make Llama available to US government agencies working on defence and national security, and to contractors including Lockheed Martin, Palantir, and Microsoft. Meta argued that it was in America's interest for American open models to succeed over models from China.

Both sides of the open-weights debate claimed the episode. Critics saw proof that open release hands powerful tools to rivals. Supporters pointed out that the model involved was an older, small one, and that open models from other countries were already freely available. Either way, it showed what an acceptable use policy can and cannot do.

From Llama to Muse (2025 to 2026)

Llama 4 did not get the reception Meta hoped for. On 30 June 2025, Mark Zuckerberg told staff in a memo that he was creating Meta Superintelligence Labs, led by Alexandr Wang, the former chief executive of Scale AI, as Meta's Chief AI Officer, with former GitHub chief executive Nat Friedman. Meta had just invested 14.3 billion dollars in Scale AI. In July 2025, The New York Times reported that leaders of the new lab had discussed moving away from open models, and that testing of the unreleased Behemoth had stopped. On 30 July, in a letter on "personal superintelligence", Zuckerberg wrote that Meta would need to be "careful about what we choose to open source". A Meta spokesperson said its position on open source was unchanged and that it expected to train "a mix of open and closed models".

The mix arrived in 2026. On 8 April 2026 the new lab released Muse Spark, its first model, to power Meta AI in the Meta AI app and on the web, with a rollout to WhatsApp, Instagram, Facebook, and Messenger to follow. Unlike Llama, its weights were not published; Meta said it hoped to open-source future versions. Then on 10 August 2026 Meta released Muse Glimmer, a 30B model distilled from Muse Spark, with downloadable weights under the Apache 2.0 licence, a standard open-source licence without Llama's special conditions. Meta said it can run on a Mac or PC with a single consumer graphics card.

At the time of writing, the Llama models remain downloadable under their licences, but no new Llama generation has followed Llama 4, and Meta's newest models carry the Muse name. Plans change quickly here, so check Meta's current announcements.

An everyday case: the solicitor's private drafts

Picture Priya, who runs a two-person legal practice. Her professional rules make her wary of pasting client letters into an online chatbot. A friend installs Ollama on her office laptop and downloads a small Llama model. Now she can ask it to tidy the wording of a draft or summarise a long letter, with nothing leaving the laptop. She soon learns its limits: it is slower and less capable than the big online services, it knows nothing after its training date, and it sometimes states wrong facts with confidence, so she never lets it touch legal citations. Before using it for paid work, she reads the licence and the acceptable use policy, and finds that a two-person firm is nowhere near the 700 million user threshold.

Try it yourself

This exercise runs a small Llama model on your own computer, using the free Ollama app. It takes 20 to 40 minutes, mostly downloading. If you can't or don't want to install software, see the alternative at the end.

What you need, at the time of writing: a Windows 10 (version 22H2 or newer) or Windows 11 PC, a Mac running macOS 14 Sonoma or later, or a Linux computer; a few gigabytes of free disk space; and ideally 8 GB of memory or more. Check Ollama's download page for current requirements, and only download from its official site.

  1. Install Ollama. Go to ollama.com, click Download, choose your system, and run the installer as you would any program. At the time of writing it is free to download and use.
  2. Open a terminal. On Windows, press the Start button, type "Terminal" or "PowerShell", and open it. On a Mac, press Command and Space, type "Terminal", and press Return. A terminal is simply a window where you type commands.
  3. Download and start a small Llama model. Type this and press Enter: > ollama run llama3.2 This fetches Llama 3.2 3B, about 2.0 GB at the time of writing, and opens a chat when it's done. If your computer is older or has little memory, use the 1B model instead, about 1.3 GB: > ollama run llama3.2:1b
  4. Chat with it. When you see a prompt, type a question and press Enter. Try something practical: > Rewrite this message so it sounds friendly but firm: "Your invoice is 30 days late. Please pay now."
  5. Test it offline. Turn off your Wi-Fi and ask another question. It still answers, because the whole model is on your computer. Turn Wi-Fi back on afterwards.
  6. Find its limits. Ask about something recent, such as "Who won the most recent football World Cup?" A model's knowledge stops at its training date, and small models also get older facts wrong. Compare its answer with a web search. Then ask a tricky reasoning question and compare with a large online chatbot. Notice where a small local model is good enough (rewording, summarising short text, brainstorming) and where it isn't.
  7. Look at the licence. Type /bye to leave the chat, then type ollama show llama3.2 --license to see the licence the model came with. Skim the section headings: the additional commercial terms, the acceptable use policy link, and the "Built with Llama" rule.
  8. Tidy up. Models take up disk space. To delete the one you downloaded, type ollama rm llama3.2 (or ollama rm llama3.2:1b).

Bigger Llama models need far more: at the time of writing, Ollama lists Llama 3.1 8B at 4.9 GB, the 70B at 43 GB, and Llama 4 Scout at 67 GB. The desktop app LM Studio is a no-terminal alternative; its own page recommends 16 GB of memory. Parameters explains why size matters so much.

If you can't install software: open WhatsApp, Instagram, Facebook, or meta.ai, where available in your country, and try the same invoice prompt with Meta AI. Be aware that Meta AI is a service, not a local model: your messages go to Meta's computers, and since 2026 Meta has been moving it from Llama to its newer Muse models. Read Meta's current AI privacy terms before sharing anything personal.

Common mistakes and misunderstandings

Thinking "open" means "no rules". Llama's licence and acceptable use policy apply to every copy, including ones you run at home. If you build a product with it, you also have credit and naming duties. Read the licence for the exact version you use, because terms differ between generations.

Thinking "open weights" means "open source". You get the model files, not the training data or the full recipe, and the licence restricts who may use it and for what. Under the Open Source Initiative's definition, Llama isn't open source. If that distinction matters to your organisation, look at models released under standard licences such as Apache 2.0.

Assuming Meta AI and Llama are the same thing. Meta AI is an app and assistant; Llama is a family of models. Meta AI used Llama models from 2024, but in 2026 Meta began moving it to Muse Spark. What you see in WhatsApp today may not be Llama at all.

Expecting a laptop model to match the big chatbots. A 3B or 8B model is far smaller than the models behind leading online services. It is useful for many everyday jobs, but it knows less, reasons less well, and makes more mistakes. Run your own tests on the tasks you care about, and see Inference for why running big models costs so much.

Believing a local model is automatically safe. Running locally protects your data from being sent away, but it doesn't make the answers correct, and a downloaded model may have been fine-tuned by a stranger to behave badly. Download only from the official Meta pages on Hugging Face or from well-known tools, and treat its answers with the same care as any AI.

Assuming the next version will be open too. Meta's plans changed in 2025 and 2026, and other companies' can too. If your business depends on an open model, keep a copy of the version and licence you rely on, and have a second option.

Forgetting where you live. Some Llama rights, such as those for the multimodal models, exclude people and companies based in the EU. Check that the licence actually covers you.

Related elements