Meta Unveils Advanced AI Models Introducing Llama 4 Series

Meta surprised the tech world this weekend by unveiling Llama 4, the latest generation of its renowned AI model series featuring enhanced abilities in processing text, images, and videos. Released on a rather unusual day for a tech giant, Saturday, Llama 4 signals Meta’s accelerated push to surpass competitors who have recently outshined prior models from the company, such as China’s DeepSeek, whose open-source offerings rivaled or even exceeded previous Llama models in some benchmarks.

The newest additions to the Llama 4 lineup include Scout, Maverick, and the forthcoming Behemoth model. According to Meta, each variant has been trained on vast repositories of unlabeled multimedia data, sharpening their ability to understand and generate visual content comprehensively. These advancements appear to be a direct response to DeepSeek’s capabilities, prompting Meta to convene specialized teams tasked with dissecting how rival firms managed to efficiently deliver powerful AI models at reduced operational costs.

Meta’s launch of the Llama 4 line clearly marks a significant juncture in AI tool development, laying the cornerstone for what appears to be an expansive future for its AI ecosystem.

Meta’s latest innovations, Scout and Maverick, are already accessible from Llama.com and via platforms like Hugging Face, granting broader developer access to these cutting-edge resources. Conversely, the significantly larger model, Behemoth, remains under rigorous training and is anticipated later. Notably, the tech titan has started incorporating the Llama 4 models into its AI-supported assistant, Meta AI, ushering this advanced capability into popular apps like Messenger, WhatsApp, and Instagram for users spread across 40 countries—though, for now, multimodal features are confined to U.S. users interacting in English.

A Closer Look at Maverick, Scout, and Behemoth’s Capabilities

Highlighting the capabilities of these new models, Maverick consists of a massive 400 billion parameters, but actively leverages only 17 billion of these through its network of 128 distinct “experts”. Parameters, in AI models, correlate closely with their effectiveness in addressing diverse problems, and Maverick stands out particularly in creativity-driven tasks such as content composition and dialogue-based applications. Impressively, Meta claims Maverick outperforms competitors like Google’s Gemini 2.0 and OpenAI’s GPT-4o in areas including multilingual interactions, extended conversational contexts, logic-heavy tasks, and certain coding examinations. However, it slightly trails behind more recent models including Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, and OpenAI’s GPT-4.5.

In contrast, Scout specializes in summarization and executing logical operations across expansive codebases, boasting an extraordinary capability to handle inputs up to 10 million tokens—equivalent to an enormous compilation of text stretching into millions of words. This notable feature enhances Scout’s appeal for applications requiring intricate understanding and meticulous summarization of substantial documents. More striking, Scout operates effectively on a single Nvidia H100 GPU, while Maverick requires the more robust Nvidia H100 DGX or hardware of equivalent performance benchmarks.

The upcoming heavyweight, Behemoth, promises even bigger performance metrics and demands proportionally robust hardware. With a total parameter size approaching two trillion—and actively employing roughly 288 billion parameters—Behemoth left behind competing models like GPT-4.5 and Claude 3.7 Sonnet across crucial STEM-focused evaluations. Nonetheless, even this advanced iteration could not surpass Google’s latest Gemini 2.5 Pro, an indication of how intensely competitive the top tier of AI performance has become.

Yet, none of these new Llama iterations focuses primarily on “reasoning models” in the sense championed by OpenAI’s o1 and o3-mini, which employ fact verification to enhance response accuracy but at a cost to immediacy. Despite this limitation, Meta assured users through company representatives that Llama 4 models will consistently deliver balanced, accurate responses and pledged continued improvements designed to prevent biases toward particular viewpoints or political inclinations.

Interestingly, these assurances arise amid a charged political climate surrounding AI, where powerful figures aligned with former President Donald Trump, such as Elon Musk and investor David Sacks, voice growing skepticism toward popular AI chatbots, suspecting them of suppressing conservative ideologies. Musk’s venture, xAI, itself has wrestled with accusations of bias, illustrating the generalized and unresolved difficulty of producing entirely neutral conversational AI models. Meta, alongside competitors like OpenAI, has accordingly begun efforts to enhance model transparency and responsiveness, notably toward politically sensitive or controversial inquiries from the public.