Meta Platforms has launched its latest artificial intelligence system, Llama 4, introducing two specialized variants: Scout and Maverick. This new generation represents a significant advancement in multimodal AI technology, capable of processing and interpreting text, images, video, and audio simultaneously.
Llama 4 Scout emphasizes efficiency, requiring only a single Nvidia H100 GPU while supporting an impressive 10-million-token context window. In benchmark testing, Scout has surpassed competing models from Google and Mistral. The Maverick variant targets sophisticated applications like coding and logical reasoning, demonstrating superior performance to GPT-4o and DeepSeek-V3 despite utilizing fewer active parameters.
Meta has also revealed plans for Llama 4 Behemoth, an ambitious model featuring 288 billion active parameters within a massive 2 trillion parameter architecture. Though still in development, early indicators suggest Behemoth could outperform industry leaders like GPT-4.5 and Claude Sonnet 3.7, particularly in STEM applications.
All Llama 4 models employ a “mixture of experts” framework, enhancing efficiency by activating only necessary parameter subsets for specific tasks. While Meta describes these models as open-source, their licensing terms have sparked controversy by restricting usage for companies with over 700 million users.
The Open Source Initiative has questioned the legitimacy of Meta’s open-source claims, noting that such restrictions contradict fundamental principles of truly open technology. These limitations have sparked debate within the AI development community about accessibility and commercial control of advanced AI systems.
