Meta has just launched two exciting new models from its Llama 4 series—Scout and Maverick. These aren’t just any AI models; they’re set to change how we approach artificial intelligence, especially in language and image processing. They’re open-source and multimodal, meaning they’re designed to handle a variety of complex tasks better than ever before.
Let’s start with Llama 4 Scout. This model is built with 17 billion active parameters and 16 experts. It’s perfect for summarizing long documents, analyzing huge sets of user data, and even interpreting code. What makes Scout stand out? Its speed and efficiency. It can handle up to 10 million tokens and works smoothly on a single GPU, which is quite impressive.
On the other hand, we have Llama 4 Maverick. It shares the same number of active parameters as Scout but ups the game with 128 experts. Maverick is your go-to for a wider range of AI applications. Whether you need help with AI assistance, conversations, coding, or multilingual support, Maverick is up for the challenge. Meta even claims it outshines models like GPT-4o and Gemini 2.0, and holds its own against DeepSeek v3.1 in programming tasks.
Both of these models are paving the way for the upcoming Llama 4 Behemoth supermodel. Behemoth is still in training, but it’s set to have a whopping 288 billion active parameters. That’s going to make it one of the most powerful foundation models out there.
Meta isn’t stopping there. They’re also planning to release Llama 4 Reasoning, a model focused on advanced inference capabilities. If you’re interested in checking out these models, they’re available on Meta’s website, Hugging Face, and cloud platforms like Azure, AWS, and Google Cloud. You can access them under commercial and research licenses.
