Meta Llama 4 Series: Scout, Maverick, Behemoth

Meta unveils the Llama 4 family, including Scout, Maverick, and Behemoth, detailing Scout as a 17-billion-parameter model with 16 Experts that can process up to ten million tokens of context on a single H100 GPU. Maverick employs a 128-expert Mixture-of-Experts architecture and, according to Meta, outperforms GPT-4o in reasoning, coding, and image understanding while using less compute than some rivals. Behemoth, still in training, reportedly exceeds two trillion parameters and benchmarks higher than GPT-4.5, Claude Sonnet 3.7, and Gemini Flash 2.0 Pro in math and science, serving as a “teacher” model to underpin smaller variants. Meta emphasizes ongoing open-access and multimodal capabilities, with Scout enabling multimedia text generation and visual analysis and integration into WhatsApp, Instagram, and Messenger. New training innovations include Early Fusion, FP8-precision, and a curriculum guiding prompts into post-training. The deployment may reshape content creation, marketing workflows, and brand communications, while raising considerations about data usage and privacy in AI training.