Meta has introduced a new set of AI models under the Llama family called Llama 4. Released on a Saturday, this collection includes four models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to Meta, these models were trained on extensive datasets comprising unlabeled text, images, and videos to enhance their visual understanding.
The recent advancements made by the Chinese AI lab DeepSeek reportedly accelerated the development of Llama 4. Their models have shown performance that either meets or surpasses Meta’s previous flagship offerings, prompting Meta to analyze how DeepSeek achieved lower operational costs for its models. Llama 4 Scout and Llama 4 Maverick are accessible through Llama.com and Meta’s partners, such as Hugging Face.
However, Llama 4 Behemoth is still undergoing training. The company has updated its AI assistant, Meta AI, to incorporate Llama 4 across various applications, including WhatsApp, Messenger, and Instagram, in 40 countries. Notably, the multimodal features are initially available only in the U.S. in English.
Developers may express concerns regarding the licensing of Llama 4. Users based in the EU are prohibited from utilizing or distributing these models, likely due to strict AI and data privacy regulations. Additionally, companies with over 700 million monthly active users need to secure a special license from Meta before deploying the models.
The Llama 4 collection is also significant for introducing a mixture of experts (MoE) architecture, enhancing computational efficiency. While Maverick has a colossal 400 billion parameters, it employs only 17 billion active parameters through its 128 experts. Scout, designed for tasks like document summarization, has a large context window, allowing it to handle tasks involving millions of words effectively.