Malaysia
Singapore
Philippines
Stay connected with BizTech Community—follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
Meta has unveiled a new set of AI models in its Llama family, dubbed Llama 4, with three key versions: Scout, Maverick, and the still-in-training Behemoth. The models were trained on vast quantities of text, images, and videos, aimed at significantly advancing multimodal AI capabilities.
Llama 4 Scout and Maverick are now available on Llama.com and through partners like Hugging Face. Meta AI, the company’s assistant integrated across WhatsApp, Messenger, and Instagram, has already been updated with Llama 4 in 40 countries, though multimodal support is currently limited to U.S. users in English. Behemoth, Meta’s most ambitious model in the set, is still under training.
These new models reflect Meta’s aggressive push to reclaim leadership in open-source AI. According to reports, competition from Chinese AI lab DeepSeek, which released high-performing and cost-effective models like R1 and V3, prompted Meta to accelerate Llama development. The new models mark Meta’s first foray into “mixture of experts” (MoE) architecture, which optimizes performance by activating only relevant components of the model for each task.
Scout features 17 billion active parameters, drawn from 109 billion total parameters and 16 experts. Its standout feature is a massive 10 million token context window, enabling it to handle extremely long documents and inputs. Scout is optimized for document summarization and reasoning over large codebases. It can run on a single Nvidia H100 GPU, offering an efficient option for developers with limited hardware.
Maverick, positioned as a general-purpose assistant, includes 400 billion total parameters and 128 experts, but uses only 17 billion active parameters at a time. Meta claims Maverick outperforms OpenAI’s GPT-4o and Google’s Gemini 2.0 on several tasks, including coding, multilingual reasoning, and long-context processing. However, it falls short of the newest models like Gemini 2.5 Pro, Claude 3.7 Sonnet, and GPT-4.5. Maverick requires more advanced hardware, such as Nvidia’s H100 DGX system.
Behemoth, the largest of the three, boasts 288 billion active parameters and nearly two trillion total parameters across 16 experts. Though still in training, Meta says Behemoth outperforms GPT-4.5 and Claude 3.7 Sonnet in math and science benchmarks. However, it doesn’t surpass Google’s Gemini 2.5 Pro.
Llama 4 models have also been tuned to be more responsive to political and socially sensitive prompts. Meta says the models are now less likely to reject “contentious” questions and are designed to offer factual, balanced responses without favoring particular viewpoints. This shift comes amid criticism from figures like Elon Musk and David Sacks, who accuse AI firms of politically biased outputs. Meta’s adjustments align with broader industry trends, where major companies are making their models more accommodating to controversial topics.
Still, Llama 4’s licensing has stirred controversy. The models are not available for users or organizations based in the EU, likely due to stringent data and AI regulations. Additionally, any company with over 700 million monthly active users must obtain special permission from Meta to use the models—a restriction that critics argue disqualifies Llama from being truly open source. In 2023, the Open Source Initiative stated that such conditions remove Llama models from the open-source category altogether.
Meta maintains that the goal is to democratize powerful AI. In a recent Instagram video, CEO Mark Zuckerberg reiterated Meta’s commitment to open-source AI: “Our goal is to build the world’s leading AI, open source it, and make it universally accessible so that everyone in the world benefits. With Llama 4, this is starting to happen.”
Despite limitations, the release is a major milestone. Developers can now access and experiment with Scout and Maverick directly on Meta AI platforms or integrate them into their applications. Meta is also preparing to release a standalone app for its AI assistant later this year, further expanding access.
Meta will discuss its roadmap at the first-ever LlamaCon AI conference on April 29. The event is expected to provide more technical details, user case studies, and possibly early insights into Behemoth’s progress.
With Llama 4, Meta signals a deepening commitment to shaping the next phase of AI development—open, powerful, and not without its share of political and ethical complications.