Llama is a family of large language models made by Meta. A large language model is a neural network trained on large amounts of text to predict the next word, so it can answer questions and write text. Meta publishes the model weights (the learned parameters) for anyone to download.
The Llama 3 family covers Llama 3, 3.1, 3.2 and 3.3. The first Llama 3 release had pre-trained and instruction-tuned models of 8B to 70B parameters (B means billion). Llama 3.1 added a 405B model; Llama 3.2 added 1B and 3B text models for phones and edge devices and 11B and 90B models that also read images; Llama 3.3 is a 70B text model. The newer Llama 4 models are natively multimodal and use a mixture-of-experts architecture, in which each input activates only part of the network. Developers fine-tune the models and run them on their own servers, on Mac, Linux or Windows, or in the cloud.
The weights are free to download under Meta's Llama Community License, a royalty-free license that requires companies with more than 700 million monthly active users to ask Meta for a separate license.