Meta Unveils Latest Llama 4 AI Models with Better Architecture and More

Meta has unveiled Llama 4, its latest family of large language models (LLMs). These models now power the Meta AI assistant across platforms such as WhatsApp, Messenger, Instagram, and the web.

The release includes two models available for immediate use—Llama 4 Scout and Llama 4 Maverick—while a more advanced version, Llama 4 Behemoth, is still under development.

Llama 4 Scout

Meta designed Llama 4 Scout as a smaller model to run efficiently, noting that it can operate using just a single Nvidia H100 GPU. The model supports a 10-million-token context window, allowing it to process large amounts of data at once. According to Meta, Scout outperforms Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as Mistral 3.1, on several widely used benchmarks. Meta also makes Scout available for download via its platform and on Hugging Face.

Llama 4 Maverick

The Llama 4 Maverick model is positioned as a more advanced alternative, competing with OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash. Meta claims that Maverick delivers comparable results in coding and reasoning tasks while using fewer active parameters — a design that could make it more efficient for certain applications. Meta specifically references DeepSeek-V3 as a point of comparison.

Llama 4 Behemoth

Still in training, Llama 4 Behemoth is Meta’s upcoming flagship model. It features 288 billion active parameters and a total of 2 trillion parameters. Although it’s not yet available, Meta CEO Mark Zuckerberg markets it to be the “highest performing base model in the world.” Early claims suggest that Behemoth will outperform models like GPT-4.5 and Claude Sonnet 3.7 on a range of STEM benchmarks.

MoE Architecture

For the Llama 4 family, Meta has adopted a Mixture of Experts (MoE) architecture. This approach dynamically activates different parts of the model based on the task at hand, which helps optimize performance while reducing resource usage. Meta says this shift has allowed it to develop scalable and efficient models.

Open Source with Conditions

Although Meta refers to Llama 4 as open-source, the licensing terms continue to raise concerns in the developer community. The current license stipulates that any commercial entity with more than 700 million monthly active users must request permission from Meta to use the models. This restriction has led the Open Source Initiative to argue that Llama does not meet the standard definition of open source.

Looking Ahead

Meta plans to reveal more about its future AI developments at LlamaCon, an upcoming event scheduled for April 29. The company is expected to provide further details on its roadmap for AI models and related products.