Meta Launches Llama 4 AI Models: A New Era for Flagship AI with Multimodal Capabilities and Licensing Changes
Meta has officially launched its new collection of AI models under the Llama 4 family, marking a significant step forward in AI development. Released on a Saturday, the new models include Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth—all designed to offer a broad visual understanding with their training on vast amounts of unlabeled text, image, and video data.
Why Llama 4? The Drive Behind Meta’s New Flagship Models
The development of Llama 4 models was reportedly accelerated after the success of open AI models from Chinese AI lab DeepSeek, which performed on par or even surpassed Meta’s previous Llama models. This competition prompted Meta to rapidly enhance its AI capabilities. Meta reportedly set up “war rooms” to analyze how DeepSeek reduced the cost of running and deploying models like R1 and V3.
Available Models: Scout, Maverick, and Behemoth
Three models are now available to the public: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Scout and Maverick can be accessed through Meta’s website and partners, including Hugging Face, while Behemoth remains in training. Meta’s AI-powered assistant, Meta AI, now integrates Llama 4 across popular apps like WhatsApp, Messenger, and Instagram in 40 countries. However, multimodal features are initially only available in the U.S. and in English.
Licensing Issues and Regional Restrictions
There are some licensing concerns with the Llama 4 models, particularly for users and companies based in the EU. Due to stringent AI and data privacy laws in the region, EU-based entities are prohibited from using or distributing Llama 4 models. Additionally, companies with over 700 million monthly active users must seek special permission from Meta to use the models, a policy that Meta can approve or deny at its discretion.
A New Architecture: Mixture of Experts (MoE)
One of the key innovations behind Llama 4 is its use of Mixture of Experts (MoE) architecture. MoE allows for more computational efficiency by breaking down tasks and assigning them to specialized “expert” models. For example, Llama 4 Maverick has a total of 400 billion parameters, but only 17 billion active parameters spread across 128 expert models. Meanwhile, Llama 4 Scout has 109 billion total parameters and 17 billion active parameters across 16 experts.
Performance Benchmarks and Model Strengths
Meta’s internal tests show that Llama 4 Maverick, with its focus on general assistant tasks like creative writing, outperforms models like OpenAI’s GPT-4 and Google’s Gemini 2.0 on coding, reasoning, multilingual tasks, long-context processing, and image benchmarks. However, it falls short compared to newer models like Google’s Gemini 2.5 Pro and OpenAI’s GPT-4.5.
Llama 4 Scout excels in document summarization and handling large codebases, and its standout feature is its massive context window of 10 million tokens, allowing it to process incredibly long documents with images and millions of words. On hardware requirements, Scout can run on a single Nvidia H100 GPU, while Maverick needs an Nvidia H100 DGX system or similar.
Meta’s unreleased Llama 4 Behemoth is expected to have 288 billion active parameters and nearly two trillion total parameters. Early benchmarks suggest that Behemoth will outperform GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in STEM-related tasks but won’t match the more advanced Google Gemini 2.5 Pro.
Llama 4’s Approach to Controversial Content
Meta has also made efforts to adjust Llama 4’s approach to handling “contentious” political and social questions. The models are designed to refuse to answer controversial queries less often, providing more balanced and factual responses. Unlike previous versions of Llama, Llama 4 is more responsive to a wider range of questions, especially on debated topics, and avoids favoritism towards specific viewpoints.
This change comes amid ongoing debates about AI chatbot bias. Some political figures, including close allies of President Donald Trump, have accused AI models of censoring conservative views. Despite these accusations, Meta insists that bias in AI remains a complex challenge, and it continues refining Llama models to respond more fairly to diverse perspectives.
Conclusion: A Major Leap for AI Development
The launch of Llama 4 marks the beginning of a new era for Meta’s AI ecosystem. With innovations like MoE architecture, vast multimodal capabilities, and more balanced responses to controversial topics, Llama 4 sets a new standard for AI models in both performance and accessibility. As Meta continues to refine these models, Llama 4 could potentially shape the future of AI-powered assistants, creative tools, and more across various industries.
