Meta has unveiled its latest generation of artificial intelligence models, part of the Llama family, releasing three new models over the weekend in a surprise announcement on Saturday, April 5. The Llama 4 suite – comprising Scout, Maverick, and Behemoth – marks a significant leap in the company’s ambition to shape the future of open AI, with models designed to tackle a range of tasks, from document summarization to advanced reasoning across text, images, and video.
The models are built on a new “mixture of experts” (MoE) architecture, which promises greater efficiency by delegating tasks to specialised components within the model. Meta claims that Maverick, its flagship general-purpose assistant, can outperform OpenAI’s GPT-4 and Google’s Gemini 2.0 on a range of coding, reasoning, and image-based benchmarks. However, it falls short of OpenAI’s most advanced GPT-4.5 and Google’s newer Gemini 2.5 Pro, according to TechCrunch.
Scout and Maverick are now freely available on Meta’s website and through partners, including the AI platform Hugging Face, though their use comes with notable caveats. Most significantly, Meta is barring companies and developers based in the European Union from using or distributing the models — a move likely driven by the region’s stringent AI governance and data privacy laws. Meta has previously criticized the EU’s regulatory framework as being heavy-handed and stifling innovation.
Open race heats up
The release follows a flurry of activity in the open-source AI world, spurred in part by the rapid ascent of Chinese lab DeepSeek. Its models — notably R1 and V3 — have performed competitively against Llama 2, prompting Meta to accelerate Llama 4 development, reportedly launching internal “war rooms” to reverse engineer DeepSeek’s efficiency gains.
Of the three new models, Scout is the most lightweight, with 17 billion active parameters and an impressive 10 million token context window. This enables it to process extensive documents and large code bases, making it suitable for academic research, enterprise data analysis, and legal work. It’s also optimised to run on a single Nvidia H100 GPU, allowing more modest deployments compared to its heavier sibling.
Maverick, meanwhile, boasts 400 billion parameters (17 billion active across 128 experts) and is designed for more general AI assistant tasks, such as creative writing and language comprehension. It requires enterprise-level compute infrastructure, including Nvidia’s DGX systems, to operate effectively.
Still in training is Behemoth, a model that Meta says will eclipse rivals on key STEM benchmarks. With 288 billion active parameters and nearly two trillion in total, it’s among the most significant AI models ever publicly described. Meta’s early tests suggest that Behemoth could surpass GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in solving complex mathematical and scientific problems — although Gemini 2.5 Pro remains ahead on several fronts.
