Meta has announced the rollout of the newest suite of artificial intelligence models called Llama 4, the latest additions to its Llama family, notably unveiled on a Saturday. The release comprises three distinct models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to Meta, each model has been trained extensively using large quantities of unlabeled textual content, imagery, and video data, granting them significant visual comprehension capabilities.
The motivation behind the rapid advancement of these models reportedly comes from the rising prominence of open-source models by Chinese AI researchers at DeepSeek. DeepSeek’s recent offerings have matched or surpassed the performance of previous top-tier Llama models, thereby accelerating Meta’s internal development processes. Sources indicate that Meta swiftly established discussion and planning groups intended to understand how DeepSeek successfully brought down computing costs required for deploying models like R1 and V3.
Currently, the Scout and Maverick models are live and accessible through the Llama.com website and via various partners, including Hugging Face, an esteemed AI development platform. The more powerful Behemoth model, however, is still undergoing its training phase. In addition, Meta has integrated Llama 4 into its own Meta AI assistant, which serves users through platforms such as WhatsApp, Instagram, and Messenger in over 40 countries. At present, the multimodal functionalities of these new models are restricted solely to English-language queries within the United States.
Despite the release’s excitement, the licensing terms surrounding the Llama 4 models may present concerns for certain developers and businesses. Specifically, the terms exclude organizations and users based or primarily operating within the European Union from utilizing or distributing these models, presumably due to strict regulatory frameworks and data privacy legislation implemented in that region. Furthermore, regarding larger firms, Meta continues its previous practice whereby enterprises with more than 700 million monthly active users must request specialized licenses, with Meta retaining complete discretion to either approve or decline these requests.
Meta has described the release of Llama 4 as a significant turning point, suggesting this new collection merely heralds the initial phase in a broad expansion of the Llama ecosystem. The latest models also introduce “mixture of experts” (MoE) architecture for the very first time, which Meta highlights as computationally more effective in both its training processes and in responding to queries and tasks. Under the MoE architecture, complex data processing jobs are segmented into smaller, manageable tasks, each of which is handled by specialized sub-models described as “experts.”
To illustrate, Maverick includes around 400 billion total parameters altogether, but only actively engages 17 billion parameters over 128 expert units. On the other hand, Scout features 109 billion total parameters, with 17 billion active parameters divided among 16 experts. Meta’s internal evaluations indicate that Maverick excels specifically in scenarios involving creativity, conversational interaction, multilingual capability, extended context management, image analysis, and certain coding and logical reasoning tasks, surpassing well-known models such as OpenAI’s GPT-4 and Google’s Gemini 2.0. However, Maverick falls short compared to newer, more potent models like Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, and the upgraded GPT-4.5 by OpenAI.
Scout, meanwhile, stands out prominently in tasks related to extensive textual summarization and navigating large-scale codebases. Particularly noteworthy is its immense context processing capability, able to handle inputs of up to 10 million tokens, meaning it effectively processes images as well as massive textual documents containing several million words, making it extremely valuable for handling extensive documentation.
Regarding hardware requirements, Scout can be operational using merely a single Nvidia H100 GPU, whereas Maverick demands the use of an Nvidia H100 DGX system or an equivalent configuration. Yet, even more demanding will be the yet-to-be-released Behemoth variant, which Meta reveals will operate through 288 billion active parameters distributed across 16 experts, reaching an almost staggering two trillion total parameters. Meta’s evaluations suggest substantial performance improvements for Behemoth across STEM-oriented competencies, outperforming models such as GPT-4.5, Claude 3.7 Sonnet, and Google’s Gemini 2.0 Pro, although still trailing slightly behind Gemini’s latest 2.5 Pro edition.
However, it is critical to mention that none of these new models belong to the category of dedicated “reasoning models,” akin to OpenAI’s recently developed o1 or o3-mini, intended to rigorously verify answers and provide greater reliability, but typically at the expense of slower response times.
Additionally, Meta has intentionally adjusted the newest Llama models to actively address more of what it terms “contentious” topics, reducing previous limitations placed on discussing social and political issues. By engaging more openly with debated subjects, Meta’s spokesperson emphasized that these models produce balanced and unbiased outcomes. Meta assures users of impartial, accurate responses from Llama 4 models, enhancing responsiveness to various viewpoints without promoting or discouraging specific perspectives.
This change arrives amid heightened scrutiny from political circles, notably those closely aligned with the White House, about perceived biases within AI chatbots and their alleged “woke” political leanings. Such accusations have prominently arisen from influential figures surrounding former President Trump, including the entrepreneur Elon Musk and AI advocate David Sacks, who have argued that popular AI assistants suppress conservative ideological content. Previously, Sacks specifically criticized OpenAI’s ChatGPT for supposedly promoting political correctness and not truthfully presenting sensitive political discussions.
Yet these assertions overlook the fundamental and complex technical challenges inherent in entirely eliminating bias from AI systems. Musk’s own AI venture, named xAI, has itself encountered significant difficulty developing chatbot models that remain politically neutral.
Despite the formidable nature of this technical dilemma, AI creators such as OpenAI have nonetheless undertaken efforts in refining their models, demonstrating an increased willingness to address complex, sometimes controversial, questions directly. Meta’s launch of the Llama 4 models aligns with this broader industry trend toward more open and inclusive AI responses.