Meta has unveiled its latest AI innovation, Llama 4, a collection of advanced models aimed at redefining artificial intelligence capabilities. The lineup includes four distinct models: Scout, Maverick, Behemoth, and L4emoth. These models leverage multimodal training on extensive datasets of untagged text, images, and videos to enhance visual comprehension. Scout and Maverick are currently accessible via Llama.com and platforms like Hugging Face, while Behemoth remains in training. Meta has integrated Llama 4 into its AI assistant across applications such as WhatsApp, Messenger, and Instagram, offering support in 40 countries, though multimodal features are limited to English speakers in the U.S.

The development of Llama 4 was accelerated by competition from Chinese AI firm DeepSeek, whose open models rivaled Meta’s earlier Llama series. Meta responded by forming specialized teams to analyze DeepSeek’s cost-efficient model architecture. Llama 4 introduces a mixture-of-experts (MoE) architecture that optimizes computational efficiency by delegating tasks to specialized expert models. Maverick boasts 400 billion parameters but activates only 17 billion across 128 experts for tasks like general assistance and creative writing. Scout excels in summarizing documents and reasoning through complex codebases with a context window of 10 million tokens.

Meta aims to expand the utility of Llama 4 for businesses, enabling AI agents to automate tasks and provide personalized services. However, licensing restrictions apply; EU-based users are barred due to regulatory concerns, and enterprises with over 700 million monthly active users require special approval for access. These limitations have sparked criticism from developers.

Mark Zuckerberg emphasized the need for substantial computational power for Llama 4's training—10 times more than its predecessor, Llama 3. Meta views this as a necessary investment to avoid falling behind competitors like OpenAI and Google.