Meta Expands AI Ambitions with Launch of Llama 4 Models

Meta has officially launched a new collection of AI models under the Llama 4 name, offering powerful tools for developers, researchers, and businesses.

The lineup includes two released models—Scout and Maverick—and one currently in development, called Behemoth.

Key Features of Llama 4 Models:

Llama 4 Scout:

Llama 4 Scout is designed to be efficient and lightweight, capable of running on a single Nvidia H100 GPU. Despite its compact size, Scout delivers impressive results in tasks like summarizing long documents and analyzing large codebases.

One of Scout’s standout features is its ability to handle extremely long inputs—millions of words at a time—thanks to a 10-million-token context window.

It also accepts image inputs. In benchmark tests, Scout has outperformed similar models such as Google’s Gemma 3 and Mistral 3.1.

Llama 4 Maverick: 

Maverick is a larger, more advanced model that rivals top-tier AI systems like OpenAI’s GPT-4o and DeepSeek-V3.

It shines in coding, complex reasoning, and STEM-related tasks.  

Maverick unique ability is to deliver high performance while using fewer active parameters, making it both effective and resource-efficient.

Llama 4 Behemoth: 

Llama 4 Behemoth is currently under development, this model is expected to to be one of the largest and most capable AI models ever created.

Meta says Behemoth will have 288 billion active parameters, and a total of 2 trillion parameters, with the potential to outperform models like GPT-4.5 and Claude Sonnet 3.7 in scientific and technical benchmarks.

All Llama 4 models use a “mixture of experts” (MoE) architecture. This design allows the model to activate only the parts needed for each task, improving speed and efficiency. 

Meta has already integrated the Llama 4 models into its AI assistant, which is now available on WhatsApp, Messenger, Instagram, and through the web.

While Meta promotes Llama 4 as open-source, the license restricts use by commercial platforms with more than 700 million users.

This has sparked criticism from the Open Source Initiative, which argues that the limitations go against open-source principles.

Scout and Maverick are available now through Llama.com and Meta’s partners, including Hugging Face. Behemoth remains under development and has not yet been released.

All models were trained on massive amounts of unlabeled text, images, and video data, giving them broad visual and language understanding.

Meta will share more about the future of Llama 4 at its upcoming LlamaCon event, scheduled for April 29.

Background

Before Llama 4, Meta released Llama 1 in 2023 and Llama 2 in July 2023, both aimed at researchers and developers.

Llama 2 gained popularity for its open-access approach and performance rivaling other major models.

These models laid the foundation for Llama 3, which improved accuracy, reasoning, and efficiency, setting the stage for the more advanced Llama 4 series.

News Gist

Meta has launched its new AI model collection, Llama 4, featuring Scout and Maverick, with Behemoth in development.

These models offer advanced performance in reasoning, coding, and processing long documents, now integrated into Meta’s platforms and available for developers.