Overview of Llama 4 Models

Meta has launched its latest collection of AI models named Llama 4, which is part of its ongoing development in the Llama ecosystem. This new suite includes three specific models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth, all designed to enhance AI capabilities.

According to Meta, these models have been trained on extensive datasets comprising unlabeled text, images, and videos, offering them a broad visual understanding.

Development Driven by Competition

The rapid progression of Llama models can be attributed to competition from DeepSeek, a Chinese AI lab that has demonstrated benchmark performances that rival or exceed Meta’s previous models. As a result, Meta reportedly established dedicated teams to analyze how DeepSeek optimized its deployment costs for models like R1 and V3.

Different Access Levels for Each Model

Llama 4 Scout and Llama 4 Maverick are accessible on Llama.com and via partners such as Hugging Face. In contrast, Behemoth remains under development. Notably, the integration of Llama 4 within Meta AI has commenced across applications like WhatsApp, Messenger, and Instagram in 40 countries, though multimodal functionalities are currently restricted to English users within the U.S.

Licensing Considerations

The licensing terms for Llama 4 have raised concerns among developers, particularly stipulations affecting organizations based in the EU. Those with a principal place of business in the EU are banned from using or distributing these models, likely due to stringent AI and data privacy laws in the region. Additionally, companies with over 700 million monthly active users must seek special licensing from Meta.

Innovative Architecture: Mixture of Experts

The latest Llama models implement a mixture of experts (MoE) architecture, enhancing their efficiency. This design allows the models to divide tasks into specialized subtasks, assigned to smaller “expert” models. For instance, Maverick boasts 400 billion total parameters, with only 17 billion actively engaged across 128 experts, while Scout contains 17 billion active parameters and 16 experts with a total of 109 billion parameters.

Performance Insights

Initial tests reveal that Maverick excels in general assistant and chat functions, outperforming well-known models like OpenAI’s GPT-4 and Google’s Gemini 2.0 on various benchmarks related to coding, reasoning, and multilingual tasks. However, it still falls short against newer models such as Gemini 2.5 Pro and Claude 3.7 Sonnet.

Meanwhile, Scout excels in summarizing documents and understanding extensive codebases, facilitated by its impressive context window of up to 10 million tokens, enabling it to process vast amounts of text and images efficiently.

Future Prospects with Behemoth

The unreleased Behemoth model is anticipated to require even more robust hardware, featuring 288 billion active parameters and nearly two trillion total parameters. Benchmarks suggest Behemoth may outperform several notable models in STEM-related evaluations.

Addressing Controversial Topics

Meta has also adjusted Llama 4’s response system, making it less likely to shy away from contentious issues, responding with greater balance on politically debated topics. A Meta spokesperson emphasized the intention for Llama 4 to provide factual responses without bias, addressing allegations from various political figures regarding perceived censorship in AI communications.