Meta Unveils Two New Llama 4 AI Models, Includes One That Can Run Efficiently on Single NVIDIA H100

Meta just launched Llama 4, bringing out two new open-source AI models called Llama 4 Scout and Llama 4 Maverick. These are part of Meta’s work to make their AI tools even better, and now they’re built into the Meta AI assistant. You can use this assistant on apps like WhatsApp, Messenger, Instagram Direct, or on the Meta AI website.

Llama 4 Scout has 17 billion active parameters and 16 experts (109 billion parameters total) and can run on just one powerful graphics card, like the NVIDIA H100. Meta says this is a big deal because it means you don’t need a ton of expensive gear to use a powerful AI model. Scout is built from the start to understand both words and images together, something Meta calls “early fusion.” They say this makes it great at figuring out pictures or mixing text and visuals, like a super-smart assistant.

With 17 billion active parameters and 128 experts (400 billion parameters total), Maverick is built for tough tasks. Meta says it matches or beats big names like GPT-4o and Gemini 2.0 Flash in thinking, coding, and picture tests, often using less energy. Like Scout, Maverick handles both images and words from the start, but Meta says it’s even better at it. It can connect what you say to specific parts of a picture (like pointing out exactly what you mean), which is perfect for fancy chatbots or helpers.