Meta released Llama 4, the latest version of their language models, featuring native multimodality, longer context windows, and a more efficient Mixture of Experts (MoE) architecture. Key features include support for up to 10 million tokens in context, the ability to handle both text and image inputs, and increased training and inference efficiency. The model family includes Llama 4 Scout, Maverick, and the upcoming Behemoth, designed for various specialized tasks.
Table of contents
Key FeaturesVariantsArchitecture: Mixture of ExpertsMultimodal by DesignLong Context WindowsLanguage SupportAccess and LicensingWhat’s Next?33 Impressions1 Comment