Meta Launches Llama 4, Enhancing AI Capabilities with New Flagship Models
Meta has unveiled its latest set of AI models—Llama 4—as part of its ongoing push into advanced artificial intelligence. Released unexpectedly on a Saturday, the new lineup includes three distinct models: Llama 4 Scout, Llama 4 Maverick, and the yet-to-be-released Llama 4 Behemoth.
Each model was trained on extensive datasets comprising unlabeled text, images, and videos, enabling them to achieve comprehensive visual and contextual understanding. The push for releasing these new models was reportedly accelerated by competition from China’s DeepSeek, whose recent open models have rivaled or surpassed Meta’s previous Llama versions.
Two of the new models, Scout and Maverick, are already accessible to the public through Llama.com and partnering platforms like Hugging Face. The Behemoth model remains under development.
Llama 4 utilizes a “mixture of experts” (MoE) architecture, significantly boosting computational efficiency by dividing data-processing tasks into specialized subtasks handled by smaller expert models. For instance, Maverick contains 400 billion total parameters but activates only 17 billion across 128 specialized experts. Scout operates with 17 billion active parameters across 16 experts from a total of 109 billion parameters.
Internal tests indicate Maverick excels in general assistant tasks, creative writing, coding, and multilingual capabilities, outperforming notable competitors like GPT-4o and Google’s Gemini 2.0 in certain areas. However, it falls short against the latest leading models such as Google’s Gemini 2.5 Pro and OpenAI’s GPT-4.5.
Scout, particularly adept at summarizing extensive documents and analyzing large codebases, features an exceptionally large context window capable of processing 10 million tokens at once. Scout can operate efficiently on a single Nvidia H100 GPU, while Maverick demands more advanced systems like the Nvidia H100 DGX.
Meta’s forthcoming Behemoth model is even more ambitious, boasting 288 billion active parameters and nearly two trillion total parameters. Preliminary benchmarking places Behemoth ahead of GPT-4.5 and Claude 3.7 Sonnet in STEM-related tasks.
However, none of these new models employ specialized reasoning architectures, unlike some advanced competitors from OpenAI.
Significantly, Meta has adjusted Llama 4 models to address “contentious” and politically sensitive questions more frequently and neutrally. The models aim for balanced and factual responses without bias, a move likely prompted by critiques from conservative figures who claim AI systems tend toward politically “woke” responses.
Meta spokespersons emphasized the company’s commitment to neutrality, aiming for Llama 4 to engage with a wide spectrum of viewpoints without favoring specific ideologies.
The release highlights Meta’s continued evolution within the AI space and signals a more open approach to discussing sensitive and debated topics.
Photo Credit: DepositPhotos.com