Meta launched its Llama 4 family of AI models, including Scout, Maverick, and Behemoth, trained on extensive text, image, and video data for enhanced visual understanding. The release was partly driven by competition from Chinese AI company DeepSeek.
Scout and Maverick are publicly available on Llama.com and through partners like Hugging Face, while Behemoth remains under development. However, EU users and companies, and those with over 700 million monthly active users, face restrictions on usage and distribution.
Llama 4 models utilize a mixture-of-experts (MoE) architecture for better efficiency. Maverick, with 400 billion total parameters, excels in tasks like creative writing, surpassing models like GPT-4o and Gemini 2.0 in some benchmarks, though falling short of newer models. Scout, possessing a vast 10-million-token context window, shines in document summarization and codebase analysis. The unreleased Behemoth surpasses even GPT-4.5 on certain STEM assessments.
Meta adjusted Llama 4 to reduce refusal to answer contentious questions and to be more balanced in its responses to various viewpoints. This addresses concerns regarding political bias in AI chatbots raised by some, including allies of President Donald Trump, who allege censorship of conservative views.