Meta has recently introduced its latest generation of large language models (LLMs), named Llama 4, marking a significant advancement in the field of artificial intelligence (AI). This new model family consists of three distinct versions: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth, each tailored for specific specifications and applications.
Llama 4 Scout is the smallest of the three models, featuring 17 billion parameters for processing requests and a total of 109 billion parameters. A key advantage of Scout is its capability to operate on a single NVIDIA GPU and its expansive context window, accommodating up to 10 million tokens—an increase of 80 times compared to its predecessor.
The mid-range model, Llama 4 Maverick, also processes requests using 17 billion parameters but boasts a higher total of 400 billion parameters, managed across 128 experts. This model is optimized for handling more complex tasks, excelling in coding and reasoning challenges.
As the largest model in the lineup, Llama 4 Behemoth is currently in training and is expected to feature 288 billion parameters for running tasks, with a total parameter count of 2,000 billion. Meta anticipates that Behemoth will surpass leading competitors such as GPT-4.5 and Claude Sonnet 3.7 in performance on science, technology, engineering, and mathematics (STEM) assessments.
Both the Scout and Maverick models utilize the “Mixture of Experts” (MoE) architecture, which enhances efficiency by activating only a subset of parameters during processing—thus optimizing resource usage and improving performance. These models also support multimodal capabilities, allowing them to process text and images simultaneously, broadening their applicability across various domains.
Meta has integrated the Llama 4 models into its AI offerings on platforms like WhatsApp, Messenger, Instagram Direct, and the Meta AI website. Furthermore, these models are accessible via cloud platforms such as Azure AI Foundry and Azure Databricks, facilitating easy deployment for developers.
The launch of Llama 4 underscores Meta’s commitment to competing with industry leaders like OpenAI and Google DeepMind in the AI landscape. The company is planning to invest up to $65 billion by 2025 to enhance its AI infrastructure and develop more advanced models.
However, the development of Llama 4 has faced challenges, with reports indicating performance issues in internal testing related to mathematical and reasoning capabilities. In response, Meta has implemented new training methodologies, including the adoption of the MoE architecture, to address these challenges.