In a bold move on an unexpected day—Saturday—Meta has officially released its latest suite of large language models under the Llama family: Llama 4. This new collection marks a significant leap forward in Meta’s AI development efforts, offering enhanced performance, a more efficient architecture, and broader capabilities across various tasks.
The Llama 4 Lineup: Scout, Maverick, and Behemoth
The Llama 4 lineup introduces three key models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Each model was trained using a vast corpus of unlabeled text, images, and videos to provide a more comprehensive understanding of multimodal data. While Scout and Maverick are already publicly available through Llama.com and platforms like Hugging Face, Behemoth is still undergoing training.
These models are Meta’s first to adopt a Mixture of Experts (MoE) architecture, designed to improve both training efficiency and real-time performance. In this setup, tasks are divided and processed by specialized sub-models, or “experts.” For instance, Maverick has a staggering 400 billion total parameters, but only 17 billion are active at any one time, spread across 128 experts. Scout, the lighter model, operates with 109 billion total parameters and 16 experts, making it significantly easier to deploy—even on a single Nvidia H100 GPU.
Behemoth, the most powerful of the trio, boasts 288 billion active parameters and a massive two trillion total parameters, requiring advanced computational resources for deployment. Meta claims that Behemoth outperforms other major models like GPT-4.5 and Claude 3.7 Sonnet in key benchmarks related to STEM tasks, particularly in mathematics and logical reasoning.
Competitive Edge and Use Cases
According to internal evaluations by Meta, Maverick delivers strong performance in general assistant applications—creative writing, multilingual understanding, reasoning, and coding. Scout, meanwhile, shines in areas like summarizing long documents and working with extensive codebases. It features an extraordinary 10-million-token context window, enabling it to handle massive input lengths across text and images.
Meta’s decision to accelerate the development of Llama 4 was partly influenced by competition from Chinese AI lab DeepSeek, whose open-source models recently achieved impressive benchmark scores and operational efficiency. Meta reportedly launched internal “war rooms” to study DeepSeek’s methods and optimize its own models in response.
Global Availability and Licensing Restrictions
Llama 4 models are currently integrated into Meta AI, the company’s AI assistant across WhatsApp, Messenger, and Instagram, now live in 40 countries. However, the full suite of multimodal features is currently restricted to English-speaking users in the United States.
In terms of licensing, Llama 4 continues to present limitations similar to previous versions. Companies or individuals located in the European Union are prohibited from using or distributing the models due to regulatory constraints under the region’s AI and data privacy laws. Additionally, companies with more than 700 million monthly active users must obtain a special license from Meta, which the company may approve or deny at its discretion.
A Shift Toward More Open Dialogue
Perhaps most controversially, Meta claims that Llama 4 has been designed to engage more openly with sensitive or debated topics. The company states that it has deliberately tuned the models to avoid refusing answers to “contentious” political and social questions. This update comes amidst growing criticism from political figures and public commentators, especially in the U.S., about perceived ideological biases in mainstream AI systems.
“Llama 4 provides helpful, factual responses without judgment,” said a Meta spokesperson. “We’re making Llama more responsive so it can answer a wider range of questions from diverse viewpoints.”
Final Thoughts
The release of Llama 4 signals Meta’s renewed commitment to competing at the forefront of AI innovation. With cutting-edge architecture, improved computational efficiency, and broader functionality, these models are poised to play a significant role in shaping the next generation of AI-powered tools and applications. However, the licensing challenges and regional restrictions may limit access for some users and businesses—at least for now.
As the AI arms race continues, all eyes are on how Meta’s Llama 4 collection will influence the broader landscape of artificial intelligence.
