Llama

Meta's family of openly released large language models, widely used as the language backbone inside other models.

Llama is Meta's family of large language models. The original LLaMA was released in February 2023, followed by Llama 2 (July 2023, the first version whose license allowed commercial use), Llama 3 / 3.1 (2024, up to 405B parameters), Llama 3.2 (September 2024, which added 11B/90B vision versions and small 1B/3B models), and Llama 4 (April 2025, which switched to a mixture-of-experts architecture and is natively multimodal). The weights can be downloaded, but the license carries usage restrictions, so ‘open-weight’ is a more accurate description than ‘open source’ in the strict sense. Many academic vision-language models and VLAs use Llama as their language backbone — OpenVLA, for instance, is built on Llama 2 7B. As of April 2026, Meta's own Meta AI assistant is reportedly powered instead by the closed-source Muse model family from its superintelligence lab.

ExampleOpenVLA combines the Llama 2 language model with a visual encoder that fuses DINOv2 and SigLIP features, trained on about 970,000 real-robot demonstrations, to produce a 7B-parameter open VLA.

Also called

LLaMA, Llama 2, Llama 3, Llama 4

Related

Large Language Model · Open-weight Model · Mixture of Experts · OpenVLA · Prismatic VLMs · Backbone Network

Sources

Wikipedia: Llama (language model)
Meta AI Blog: The Llama 4 herd
OpenVLA: An Open-Source Vision-Language-Action Model (arXiv:2406.09246)

As of

2026-09