Meta Unveils Llama 4 AI Models with Advanced Multimodal Capabilities

Meta Platforms, Inc. has introduced Llama 4, a new series of AI models designed to enhance multimodal experiences, as announced on their website. The lineup includes Llama 4 Scout, Maverick, and Behemoth, each offering unique capabilities.

Llama 4 Scout and Maverick are openly available for download on llama.com and through partners like Hugging Face. Scout is optimized for tasks such as document summarization and reasoning over large codebases, featuring a context window of 10 million tokens. Maverick, with 400 billion total parameters, excels in general assistant and chat use cases, outperforming models like OpenAI's GPT-4o on certain benchmarks.

The Behemoth model, still in training, is expected to require more robust hardware and aims to surpass existing models in STEM-related tasks. Notably, the Llama 4 models employ a mixture of experts (MoE) architecture, enhancing computational efficiency by delegating tasks to specialized expert models.

However, the use of Llama 4 models is restricted in the EU due to regional AI and data privacy laws, and companies with over 700 million monthly active users must seek special licensing from Meta. These models are part of Meta's ongoing efforts to provide advanced AI solutions while addressing concerns about bias and responsiveness in AI systems.