Meta Unveils Llama 4: Multimodal AI Models with MoE Architecture and Expanded Capabilities

Meta has officially launched Llama 4, the latest generation in its open-source AI model lineup. This new collection includes three models: Llama 4 Scout, Maverick, and Behemoth. Each was trained on vast quantities of text, image, and video data, emphasizing a strong multimodal understanding. Notably, Llama 4 uses a Mixture of Experts (MoE) architecture to improve computational efficiency. Maverick, designed for general assistant tasks, features 400 billion parameters but activates only 17 billion per query, while Scout, optimized for summarization and code reasoning, supports an impressive 10 million token context window and runs on a single Nvidia H100 GPU. Behemoth, still in development, boasts nearly two trillion parameters and aims to outperform leading models in STEM tasks.

Meta claims Llama 4 outperforms GPT-4 and Gemini 2.0 in some areas, though it trails newer models like GPT-4.5 and Gemini 2.5 Pro. The models are partly available via Llama.com and Hugging Face, but access is restricted for companies in the EU and those with over 700 million users without special licensing. The new models are also tuned to engage more openly with political and social topics. Meta updated its AI assistant across apps with Llama 4 in 40 countries, though multimodal features are currently U.S.-only.