Meta releases Llama 4, a new crop of flagship AI models

Meta has introduced a new set of AI models, Llama 4, within its Llama family, launching them on a Saturday.

The collection includes four new models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Meta states that they were trained on a vast amount of unlabeled text, image, and video data to enhance their visual understanding.

The development of open models by Chinese AI lab DeepSeek, which reportedly outperformed Meta’s previous Llama models, accelerated the progress of Llama development. Meta responded by analyzing how DeepSeek reduced the cost of running and deploying models like R1 and V3.

Scout and Maverick are now openly accessible on Llama.com and through Meta’s partners, including Hugging Face, an AI development platform. Behemoth is still undergoing training. Meta has updated its AI-powered assistant, Meta AI, across various apps to utilize Llama 4 models in 40 countries, with multimodal features currently limited to the U.S. in English.

Certain developers may have concerns about the Llama 4 license, particularly its restrictions on users and companies based in the EU and companies with over 700 million monthly active users.

Meta describes the release of Llama 4 models as the start of a new era for the Llama ecosystem, with more to come in the future.

Meta reveals that Llama 4 models utilize a mixture of experts (MoE) architecture, enhancing computational efficiency for training and answering queries. MoE architectures divide data processing tasks into subtasks and assign them to specialized “expert” models. 

Maverick, for instance, contains 400 billion total parameters, with only 17 billion active parameters distributed across 128 “experts.” Scout, on the other hand, has 17 billion active parameters, 16 experts, and a total of 109 billion parameters.

Meta’s internal testing shows that Maverick excels in “general assistant and chat” applications like creative writing, surpassing models such as OpenAI’s GPT-4o and Google’s Gemini 2.0 in various benchmarks. However, Maverick falls short compared to more recent models like Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, and OpenAI’s GPT-4.5.

Scout’s strengths lie in tasks such as document summarization and reasoning over extensive codebases. It boasts a large context window of 10 million tokens, enabling it to process and work with lengthy documents and images.

While Scout can operate on a single Nvidia H100 GPU, Maverick requires a more advanced Nvidia H100 DGX system or equivalent, according to Meta’s calculations.

Behemoth, which is yet to be released, demands even more robust hardware. With 288 billion active parameters, 16 experts, and nearly two trillion total parameters, Behemoth outperforms several models in STEM skills evaluations.

None of the Llama 4 models are classified as “reasoning” models like OpenAI’s o1 and o3-mini, which prioritize fact-checking and reliability in responses but may take longer to deliver answers.

Meta has fine-tuned its Llama 4 models to handle “contentious” questions more judiciously. The models now address debated political and social topics and are more discerning about the prompts they respond to.

Meta aims to make Llama 4 more responsive across various viewpoints and ensure unbiased responses to questions. This initiative comes in response to criticisms of AI chatbots being politically biased.

Despite efforts to reduce bias in AI, it remains a complex technical challenge. Companies like OpenAI are adjusting their models to tackle controversial subjects and provide more comprehensive answers.