Llama is Meta’s family of open-weight large language models, which the company describes as open models that users can download, fine-tune, distill, and deploy anywhere. The line began in 2023 and has progressed through multiple generations.
As of June 2026, the Llama site presents Llama 4 with natively multimodal variants named Maverick and Scout, alongside the earlier Llama 3 series. It lists Llama 3.3 at 70B, Llama 3.2 in 1B and 3B lightweight sizes plus 11B and 90B multimodal variants, and Llama 3.1 in 8B, 70B, and 405B sizes. The page emphasizes direct downloads rather than stating specific release dates.
Distribution is open weights: the models are downloadable for self-hosting and customization, which sets Llama apart from API-only families.
Why business readers should care: Llama is the most prominent open-weight option, letting organizations run capable AI on their own infrastructure for data-control, cost, and customization reasons that hosted APIs cannot always satisfy.