Glossary
What is Llama 4
In short
Llama is a family of language models with open weights, released by Meta.
Models and products · Llama
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a big box of building parts that a factory gave to anyone who wants it, with instructions: build it, paint it, make your own thing. But the box comes with a sheet of rules that says how you may use it and how you may not. There is a smart program like that, which understands words and writes answers. You can take it for yourself, but first the grown-ups read the rules with you.
In plain words
It is a set of language models released by Meta. The models hold a dialog, summarize and write text, and the fourth generation also understands images. The files are available for download, so programmers run a model on their own computers and fine-tune it for their tasks. The community has built many improved versions on top of these models. For example, a developer takes a small model and teaches it to answer in the style of the help desk of their service.
For business
For a business in Russia, Llama has two sides. The weights are available under the Llama 4 Community License. It allows you to use, change and share the models, but when you share them it requires the label “Built with Llama” and a derived model name that starts with the word Llama. Companies whose services had more than 700 million monthly users on the release date of the fourth generation need a separate permission. The rights holder is Meta, so before launch a lawyer checks the risks of use and of public labeling. Technically, the model runs on your servers and data does not leave. Whether to take it or another open model is decided after a test on your tasks in a pilot.
For a specialist
The first generation is described in the paper by Touvron et al., “LLaMA: Open and Efficient Foundation Language Models” (February 2023): models from 7 to 65 billion parameters trained only on public datasets. The fourth generation, presented on 5 April 2025, is built on a Mixture of Experts (MoE) and is multimodal from the start: one model handles text and image tokens. Llama 4 Scout has 109 billion parameters, with 17 billion active and 16 experts, and a context window of up to 10 million tokens. Maverick has 400 billion, with 17 billion active and 128 experts. Both are available for download, and Behemoth was still in training at the time of the announcement. The Llama 4 Community License has applied since 5 April 2025 and includes Meta’s Acceptable Use Policy by reference.
Example
A development team downloads a small version of the model and checks how it answers questions about internal rules. At the same time, the company’s lawyer studies the license and decides whether this option is allowed.
Related terms
Questions and answers
Can I run Llama on my own server?
Sources
- Touvron et al., LLaMA: Open and Efficient Foundation Language Models, 27 February 2023arxiv.org
- Meta, The Llama 4 herd, 05.04.2025ai.meta.com
- Meta, Llama 4 Community License Agreement, 5 April 2025, checked 16 October 2026llama.com
- Pravo.ru, Meta designated an extremist organization (ruling of the Tverskoy District Court of Moscow), 21 March 2022pravo.ru
Not sure whether your company needs this?
AI solutions map