Meta Llama 3.2 Review — First Multimodal Llama, Built for the Edge

Meta Llama 3.2 launched September 25, 2024 at Meta Connect, introducing four models: 1B and 3B text-only models built for on-device deployment, and 11B and 90B vision models — Meta's first multimodal open-weight release. All four share a 128K context window. The vision models use a frozen-backbone cross-attention architecture, with an EU licensing restriction on the multimodal variants. This review covers the full family, vision architecture, benchmarks, hardware requirements, and on-device deployment.