Meta has quietly launched its latest generation of open-source AI models, Llama 4 and did so on a Saturday, breaking the usual weekday release pattern. This new family of models introduces significant upgrades across performance, accessibility. And political responsiveness, marking a new milestone for Meta’s AI roadmap.
The Llama 4 lineup currently includes three models: Scout, Maverick, and Behemoth. Each was trained using vast datasets made up of unlabeled text, images, and video to develop broad multimodal understanding. Meta says this multimodal foundation is a major step forward for its assistant, Meta AI, which now integrates Llama 4 across platforms like Instagram, WhatsApp, and Messenger in 40 countries. Though full multimodal functionality is, for now, limited to English-speaking users in the U.S.
Industry insiders say Meta accelerated its Llama 4 development in response to the rise of open-source Chinese models from DeepSeek. Which managed to match or outperform earlier Llama versions. DeepSeek’s efficient model architectures, especially R1 and V3. Reportedly triggered an internal scramble at Meta to understand how to cut costs and enhance scalability. The result? An upgraded Llama line with smarter training methods and more powerful capabilities.
The Llama 4 collection is also Meta’s first to adopt the Mixture of Experts (MoE) model architecture. This design splits tasks into sub-tasks handled by specialized “expert” models, improving training efficiency and inference speed. Each model in the collection reflects this strategy in different ways.
Maverick is the largest publicly available model of the three, boasting 400 billion total parameters. However, only 17 billion are active at any given time, spread across 128 expert systems. Meta says Maverick shines in general assistant tasks like creative writing, multilingual queries, and coding challenges. It even outperforms OpenAI’s GPT-4o and Google Gemini 2.0 on certain long-context and reasoning benchmarks. Though it falls just short of top-tier models like Claude 3.7 Sonnet and GPT-4.5.
Meanwhile, Scout is lighter and highly efficient. With 109 billion total parameters and 17 billion active across 16 experts. It can run on a single Nvidia H100 GPU—ideal for developers with limited hardware. Scout’s standout feature is its massive 10 million token context window. Allowing it to analyze long documents, large codebases, and even combined image-text inputs. Meta says Scout is especially strong at summarization and complex reasoning tasks.
Both Scout and Maverick are now available for developers through Llama.com and partnerships with platforms like Hugging Face.
Still in development, Behemoth is designed to push boundaries further. With nearly 2 trillion parameters and 288 billion active across 16 experts. This model is engineered for large-scale STEM reasoning tasks. Meta claims Behemoth already outperforms GPT-4.5 and Claude 3.7 on math and science benchmarks, although Gemini 2.5 Pro still holds the edge.
Given its size, Behemoth will require far more powerful hardware than Scout or Maverick. Likely a full-scale H100 DGX system or equivalent.
While the models are openly released, there are licensing caveats. Companies operating in the European Union are barred from using or distributing Llama 4 due to Meta’s interpretation of strict regional AI governance laws. This restriction is likely a response to the EU’s data protection and AI regulations. Which Meta has previously criticized as overly complex.
Furthermore, companies with over 700 million monthly active users must request a separate license from Meta—a gatekeeping policy unchanged from earlier Llama releases.
One of the most noticeable changes in Llama 4 is how it handles contentious political and social questions. According to Meta, these new models are more open to responding to complex or debated topics, addressing previous criticism that AI assistants refused too many prompts deemed controversial.
Meta says Llama 4 aims to be more balanced, providing factual responses without judgment, and ensuring the model doesn’t disproportionately reject viewpoints from any side. This move appears to answer critics—including some political figures and tech influencers—who accused earlier AI models, particularly OpenAI’s ChatGPT, of being biased or “woke.”
By tuning its models to allow more diverse responses while still maintaining factual accuracy, Meta is stepping directly into the ongoing debate over AI bias and free speech—an area where few AI companies have succeeded without controversy.
Meta is calling this release just the beginning. With Scout and Maverick now in the hands of developers and Behemoth on the horizon, Llama 4 may set a new bar for open-source multimodal models. Whether this strategy will help Meta leap ahead in the AI arms race—especially against OpenAI, Google, and Anthropic—remains to be seen.
