Meta Unveils Llama 4 Herd: The Next Leap in Multimodal AI

Meta has just introduced Llama 4, a new set of artificial intelligence models designed to handle text and images at once. Unlike previous versions, these models use a “mixture-of-experts” approach, so they’re more efficient and can tackle a wider range of tasks—like writing code, summarizing long documents, or explaining pictures. Two smaller versions, called Llama 4 Scout and Llama 4 Maverick, each have 17 billion active parameters (the “brain cells” of an AI model), but they punch well above their weight in performance.

Llama 4 Scout is built for super-long tasks and can handle up to 10 million tokens of text, which is great if you need to digest huge libraries of information. Meanwhile, Llama 4 Maverick focuses on balancing power and cost. Even though both are smaller than some competitor models, they’re trained using a “teacher” model called Llama 4 Behemoth—a much bigger system that sets a high bar for quality on math and coding problems. By learning from Behemoth, Scout and Maverick have become surprisingly strong in many areas.

Meta is making these new AI models open and easier to integrate, letting developers and companies customize them for their own needs. The models also come with safety features, like filtering tools to detect risky inputs or outputs. If you want to try them, you can download Llama 4 Scout and Llama 4 Maverick from Meta’s website or on Hugging Face. The hope is that the open access and specialized design will help people build more innovative apps, chatbots, and solutions for everyday use.