By training adapter weights without updating the language-model parameters, Llama 3.2 11B & 90B retain their text-only performance while outperforming on image understanding tasks vs closed models. Enabling developers to use these new models as drop-in replacements for Llama 3.1. pic.twitter.com/TQo4IIyqRH
— AI at Meta (@AIatMeta) 25 septembre 2024
By training adapter weights without updating the language-model parameters, Llama 3.2 11B & 90B retain their text-only performance while outperforming on image understanding tasks vs closed models. Enabling developers to use these new models as drop-in replacements for Llama 3.1.