Llama 3.2

Llama 3.2 extends Meta's Llama family in two directions: lightweight text only models at 1B and 3B parameters aimed at on-device and edge deployment, and multimodal vision models at 11B and 90B that add image understanding and visual reasoning. All variants carry a 128K token context length and TikToken based tokenization, and quantized options including FP8 and Int4 reduce memory requirements for phones and other constrained hardware. The meta-llama/llama-models repository hosts the tokenizers, CLI tools, download flow, and inference examples for the family. Distribution follows the Llama 3.2 Community License, which allows research and commercial use after license acceptance, with weights also available through Hugging Face and major cloud platforms. Mobile and edge developers reach for the small models to power local assistants, while the vision variants handle document understanding, captioning, and other image plus text workloads.