What is Llama? Llama is a family of large language models from Meta, the parent company of Facebook, Instagram and WhatsApp, which you can download free of charge and run on your own hardware. The name stands for Large Language Model Meta AI. Llama AI and Meta Llama refer to the same model family.
Llama is the best-known open-weight family from a big tech company. Since April 2026, Meta has also had a closed flagship, Muse Spark, but Llama 4 remains available.
Which versions of Llama are there?
Llama has four generations, of which Llama 4 is the most recent. Llama 1 (then written LLaMA) appeared in February 2023 and was intended for research. Llama 2 followed in July 2023 with a licence for commercial use.
Llama 3 appeared in 2024 with larger context windows and better reasoning performance. Llama 4 came out on 5 April 2025, in two published models:
- Llama 4 Scout: 17 billion active parameters (109 billion in total) and a context window of 10 million tokens.
- Llama 4 Maverick: 17 billion active parameters (400 billion in total) and a context window of 1 million tokens.
Both Llama 4 models use a mixture-of-experts architecture: only part of the model is active for each query. As a result, the hardware requirements are lower than for a traditional model of similar size, in which every parameter is involved in every query.
Llama 4 Behemoth, described at the announcement as a training model of around 2 trillion parameters, was never released publicly. As of September 2026, Meta’s Hugging Face page lists no Llama models newer than Scout and Maverick. Llama 3.3 70B from December 2024 remains available as a smaller, text-only model.
When Llama 4 was released, Meta submitted an experimental chat version of Maverick to the benchmark platform LMArena that differed from the downloadable version. After that controversy, it pays to check which model variant was tested whenever you see a benchmark claim about Llama.
Has Llama been replaced by Muse Spark?
Llama has not been replaced by Muse Spark, but it has been overtaken. Muse Spark appeared on 8 April 2026 as the first model from Meta Superintelligence Labs, the AI division Meta set up in 2025, and has since been the model behind the Meta AI app and meta.ai. On 12 October 2026, Meta introduced Muse Wonder and Muse Wonder Full, which Meta calls its most capable models to date; Wonder is available as a preview through the Meta Model API and Muse Code, and Wonder Full only to Muse Code subscribers.
Muse Spark differs from Llama on one point of principle: the model is closed. You cannot download its model weights; instead, you use Muse Spark through meta.ai, the Meta AI app or the Meta Model API, which has been in public preview since July 2026. Meta AI, the assistant in WhatsApp and Instagram, previously ran on Llama and is being moved to Muse Spark.
According to Meta, Muse Spark matches the capabilities of Llama 4 Maverick with less than a tenth of the computing power. Meta still releases open-weight models, just no longer under the Llama name: Muse Glimmer, released on 10 August 2026, has 30 billion parameters and, according to Meta, runs on a single GPU. On 15 October 2026, Muse Spark Open followed, a separate open-weight model of 816 billion parameters (42 billion active) under the Muse Model License; Muse Spark itself remains closed.
How does Llama differ from ChatGPT, Claude and Gemini?
Llama differs from ChatGPT, Claude and Gemini in three ways: the distribution model, control over the model and the effort needed to get started. You use ChatGPT, Claude and Gemini through the provider’s browser interface, app or API. You download Llama from Hugging Face and run it on your own infrastructure, so your data stays within your own environment and you pay no fee per processed token.
With closed models, the provider is in control. Anthropic decides when Claude is updated, OpenAI sets the policy for ChatGPT and Google does the same with Gemini. With Llama, you decide which version you run, when you upgrade and which changes you make.
You can only meet strict requirements on data location, version stability or fine-tuning on your own material with an open-weight model. Meta is not alone in offering open weights: Google’s Gemma follows the same principle in a lighter form.
Llama does take more work. A closed model works once you log in; an open-weight model needs hardware, installation, monitoring and maintenance. Anthropic’s Claude and Google Gemini deliver all that ready-made, with their own chat environment and API.
Open-weight is not the same as open source
Llama is open-weight, not open source: you get the trained model weights, but not the complete training data or the training code. Its use also falls under Meta’s own licence rather than a standard open-source licence such as Apache 2.0 or MIT. That puts Llama between the two camps in the distinction between open source and closed AI models.
The Llama 4 Community License allows commercial use, with a few conditions. Organisations with more than 700 million monthly active users must request a separate licence; in practice, that clause only affects big technology companies. The licence does require the notice “Built with Llama” for commercial use, and that the names of your own models built on Llama start with Llama.
European users face an extra restriction: the Llama 4 Acceptable Use Policy grants no rights to the multimodal Llama 4 models to individuals domiciled in the EU or companies with their principal place of business in the EU. Scout and Maverick are themselves multimodal, so using them only for text is no way round this for an organisation based in the EU. End users of a product that contains such a model are not covered by this restriction.
Within the Llama family, that leaves the text model Llama 3.3 70B for a European organisation. The Llama 3.3 use policy only restricts EU organisations for multimodal models, and Llama 3.3 70B handles text only.
If you are looking for a more recent open-weight model, you will need to look at alternatives under a different licence. Meta’s Muse Glimmer and Google’s Gemma 4 fall under Apache 2.0, a licence with no separate restrictions for users in the EU. Muse Spark Open falls under the Muse Model License, with Meta Platforms Ireland as licensor in the EU and a mandatory “Built with Muse Spark Open” notice. Mistral Large 4 has open weights under Mistral’s own licence, a modified MIT licence.
How do you use Llama in practice?
To use Llama in practice, you choose from three routes: self-hosting on a server, a managed cloud provider or a local program on your own computer.
To self-host, you download the Llama model weights from Hugging Face and run the model on your own hardware or a cloud server. According to Meta, Llama 4 Scout runs with quantisation on a single NVIDIA H100 GPU; Maverick needs a server with several H100 GPUs. Quantisation means storing the numbers in the model at lower precision, so it needs less memory. For an organisation with its principal place of business in the EU, the licence rules out Scout and Maverick; self-hosting there means Llama 3.3 70B or one of the open alternatives from the previous section.
Self-hosting requires knowledge of machine learning infrastructure, because you handle serving, monitoring and updates yourself. The old address llama.com now redirects to Meta’s developer portal, where Muse takes centre stage.
With a managed cloud provider, you use Llama as a service. AWS, Databricks, Google Cloud and inference providers such as Groq and Together AI offer Llama models with payment per token. According to Meta’s terms of use, the EU restriction on the multimodal Llama 4 models does not apply to end users of a product or service that incorporates such a model. If you build an application on Llama 4 yourself, even through one of these providers, you are more likely to count as a developer than an end user; if in doubt, get legal advice. You keep control over the model version and can take the weights to another platform if you wish.
You can run Llama locally with a program such as Ollama, which installs and runs open-weight models like Llama on a laptop or desktop. Llama 4 is too heavy for an ordinary laptop, but the smaller Llama 3 models do run on a computer with enough memory.
For experiments, running locally with Ollama is enough. For production without your own ML team, a managed cloud provider is more practical than self-hosting.
Who is Llama a good choice for?
Llama is a good choice for organisations with a specific combination of requirements: data that stays within their own environment, control over the model version, technical staff to manage the model, and a volume at which token costs exceed the cost of self-hosting.
Examples of such applications are high-volume customer service systems, internal knowledge bases that must not go to an external provider and products where fine-tuning on your own specialist vocabulary adds value.
For an organisation based in the EU, that choice means Llama 3.3 70B, not Llama 4. If you need a newer model generation, Gemma 4, Muse Glimmer, Muse Spark Open and Mistral Large 4 are the open-weight alternatives without that EU restriction.
For most knowledge workers, Llama is not the first choice. If you write texts, summarise documents or draft emails, you get usable results faster in the chat environment of ChatGPT, Claude or Gemini.
Llama is often mentioned in the same breath as other open-weight models such as Mistral and DeepSeek. Each of these families has its own strengths and licence terms, so compare them on licence, hardware requirements and language support before you choose.
Frequently asked questions about Llama
If you want to try Llama, the first questions are often practical ones about language support and access.
Which languages does Llama support?
Llama 4 officially supports twelve languages, according to Meta’s model card; Dutch, for example, is not among them. According to the same model card, Llama 4 was trained on a broader set of 200 languages, but for other languages the responsibility for safe use lies with the developer. Meta allows you to fine-tune the model for other languages, provided you follow the licence and the use policy. So if you work in a language outside the official twelve, test the output on your own tasks before you deploy Llama.
Is Ollama the same as Llama?
No, Ollama is not the same as Llama: Ollama is a separate program for running open models locally, while Llama is a model family from Meta. Ollama is open source under the MIT licence and is released by Ollama itself, not by Meta. Besides Llama, you can use it to install Gemma, DeepSeek, Qwen and gpt-oss, among others. Expect a hefty download with Llama 4: the smallest variant in the Ollama library is 67 GB.
What do you need to download Llama?
To download Llama through Hugging Face, you need an account and must fill in an access form from Meta before you get the model files. Meta asks for details including your full legal name and date of birth, and you agree to the Llama 4 Community License. Your data is processed under Meta’s privacy policy. Check the form carefully, because you cannot change it once submitted.
Using Llama professionally in your organisation
Using Llama professionally in your organisation takes more than a download and an installation. The team needs to set down who chooses the model version, who keeps track of the licence terms, how prompts are standardised and which checks the output passes before it leaves the system.
Meta’s change of course, exactly a year after Llama 4, shows why those agreements should not depend on a single model. A way of working tailored only to Llama 4 becomes outdated as soon as the vendor changes direction. Agreements on checks, ownership and source verification stay valid with any open-weight or closed model.
In the AI Literacy Course, you learn to set down those agreements in an AI policy with fixed checkpoints, so your team uses Llama, Muse or any other model with the same way of working. The output remains your responsibility, whichever model generates it.
