Meta has officially launched Llama 4, its latest suite of advanced artificial intelligence models designed to raise the bar in performance, efficiency, and versatility. This new generation includes three key models: Scout, Maverick, and the upcoming Behemoth, each crafted to tackle different challenges in the AI space.
Built on a mixture of experts architecture, Llama 4 models handle tasks by activating only the most relevant specialized sub-models, making them more efficient than traditional AI systems. This technique allows for more intelligent task delegation, faster processing, and lower resource consumption.
Maverick, one of the flagship models, is optimized for creative writing, general-purpose chat, and language-related tasks. It contains 400 billion parameters in total but uses only 17 billion at a time across 128 expert pathways. Internal tests from Meta suggest that Maverick surpasses OpenAI’s GPT-4o and Google Gemini 2.0 in several benchmarks, although it does fall short of the most cutting-edge systems like Gemini 2.5 Pro and GPT-4.5.
Scout is engineered for large-scale document analysis, summarization, and complex code reasoning. It supports a context window of 10 million tokens, enabling it to process extremely lengthy inputs. Scout is also lightweight enough to operate on a single Nvidia H100 GPU, making it highly accessible for developers.
Related Post
Behemoth, still under development, is Meta’s most powerful model to date. With 288 billion active parameters and a total nearing two trillion, it is designed to excel in science, technology, engineering, and math applications. Meta reports that Behemoth outperforms GPT-4.5 and Claude 3.7 Sonnet in several technical evaluations.
These models were trained using vast amounts of unstructured text, image, and video data, giving them a broad and deep understanding across multiple formats. Scout and Maverick are already available through Llama.com and Meta’s partners such as Hugging Face, while Behemoth will be released after further refinement.
Meta’s push to launch Llama 4 was partly driven by competition. The rise of open-source models from Chinese AI lab DeepSeek, which matched or exceeded the capabilities of Meta’s older models, led to an accelerated development timeline.
Llama 4 is now integrated into Meta AI, the virtual assistant available in WhatsApp, Messenger, and Instagram across 40 countries. For now, advanced multimodal features like image processing are limited to users in the United States who interact with the assistant in English.
However, usage comes with conditions. Entities based in the European Union are restricted from using or distributing the models, likely due to the region’s strict AI governance and data privacy rules. Additionally, companies with more than 700 million monthly users must obtain special permission from Meta.
Meta has also reworked how Llama 4 handles sensitive and controversial topics. Unlike previous models that often avoided such questions, Llama 4 is more open to addressing debated social and political issues while maintaining a neutral and factual tone.
This update arrives amid growing criticism over political bias in AI systems. Figures like Elon Musk and David Sacks have publicly accused leading chatbots of favoring certain viewpoints. Meta aims to position Llama 4 as a more balanced alternative that avoids promoting specific ideologies.
While complete neutrality in AI remains a complex technical goal, Meta believes that Llama 4 makes significant progress. With this release, the company is signaling its intent to remain a leader in the increasingly competitive and rapidly evolving AI industry.
