Meta surprised the tech world on Saturday, April 5, by launching its next generation of AI models in the Llama family. The company released three models—Scout, Maverick, and Behemoth—as part of the new Llama 4 suite.

New Models, New Capabilities

The models handle tasks from document summarisation to advanced reasoning across text, images, and video. They mark a significant step in Meta’s goal to shape the future of open AI. Meta built the new models on a “mixture of experts” (MoE) architecture. This design improves efficiency by letting specialised components handle different tasks. Meta claims that Maverick, its flagship assistant, can outperform OpenAI’s GPT-4o and Google’s Gemini 2.0 on coding, reasoning, and image-based tests. However, it falls short of OpenAI’s GPT-4.5 and Google’s Gemini 2.5 Pro, TechCrunch reported.

Availability and Restrictions

Scout and Maverick are now available on Meta’s website and through partners like Hugging Face. Their use, however, comes with restrictions. Meta bars companies and developers in the European Union from using or distributing the models. This step likely responds to the EU’s strict AI governance and data privacy laws. Meta has called the EU framework heavy-handed and innovation-stifling.

Open-Source Competition

The launch comes amid a burst of activity in the open-source AI world. Chinese lab DeepSeek gained attention with models R1 and V3. These models performed well against Llama 2. In response, Meta sped up Llama 4 development and set up internal “war rooms” to study DeepSeek’s efficiency gains.

Scout is the lightest model. It has 17 billion active parameters and a 10 million-token context window. This setup lets it process long documents and codebases. It is optimised to run on a single Nvidia H100 GPU.

Maverick targets general AI tasks like creative writing and language comprehension. It features 400 billion parameters, with 17 billion active across 128 experts. It requires enterprise-level compute power, such as Nvidia’s DGX systems.

Behemoth is still in training. Meta says it will excel in key STEM benchmarks. It boasts 288 billion active parameters and nearly two trillion in total. Early tests suggest that Behemoth could beat GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in complex math and science tasks. However, Gemini 2.5 Pro still leads in some areas.

Meta’s launch of the Llama 4 suite marks a major leap forward in AI. The new models show the company’s strong commitment to advancing open AI.