Meta presents Llama 4, a new generation of prominent artificial intelligence models.

Meta has introduced a new series of artificial intelligence models called Llama 4, which joins the Llama family in a surprisingly launched release on a Saturday. The collection includes four models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to the company, these models were trained using “large amounts of unlabeled text, image, and video data,” granting them a “broad visual understanding.” The development of Llama 4 has been accelerated, in part, by the success of open models from the Chinese lab DeepSeek, which have demonstrated comparable or superior performance to earlier versions of Llama. This has led Meta to organize working groups to analyze how DeepSeek has reduced the costs of running and deploying models such as R1 and V3.

Scout and Maverick are openly available at Llama.com and through Meta partners, including the AI development platform Hugging Face, while Behemoth is still in the training process. Meta has reported that its AI assistant, Meta AI, which operates on applications like WhatsApp, Messenger, and Instagram, has been updated to use Llama 4 in 40 countries, although multimodal features are currently limited to the USA and in English.

One of the controversies surrounding Llama 4 is its usage license, which prohibits users and companies based in or having a “principal place of business” in the EU from using or distributing the models, likely due to regulations imposed by data privacy and governance laws in the region. Additionally, as with previous Llama releases, companies with more than 700 million monthly active users must apply for a special license from Meta, which has the authority to approve or deny such requests at its sole discretion.

Meta emphasizes that Llama 4 is the first group of models to utilize a mixture of experts (MoE) architecture, enhancing computational efficiency in training and query responses. This architecture divides data processing tasks into subtasks assigned to smaller, specialized “expert” models. For example, Maverick has a total of 400 billion parameters, but only 17 billion are active, 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 suggests that Maverick, deemed optimal for general assistance and chat tasks, outperforms models like OpenAI's GPT-4 and Google's Gemini 2.0 in certain coding, reasoning, and image processing benchmarks.

However, Maverick does not compare to more advanced models like Gemini 2.5 Pro and Claude 3.7 Sonnet. Scout excels in document summarization tasks and reasoning over large codebases, boasting a notably wide context window of 10 million tokens, allowing it to process lengthy documents. In terms of requirements, Scout can operate on a single Nvidia H100 GPU, while Maverick needs an Nvidia H100 DGX system or equivalent, and Behemoth will require even more robust hardware.

As for Behemoth, it has yet to be publicly launched, but it reportedly has 288 billion active parameters and nearly two trillion total parameters, with performance surpassing GPT-4.5 and Claude 3.7 Sonnet in STEM-related skills evaluations. It is relevant to mention that none of the Llama 4 models are “reasoning” models like OpenAI's o1 and o3-mini.

Meta has adjusted Llama 4 to make the models less likely to answer “controversial” questions, addressing political and social topics that previous models avoided. The company claims that Llama 4 will provide useful and objective responses without making judgments. This comes at a time when AI chatbots have been criticized for being “too progressive” in their responses, a critique that has arisen from certain sectors close to the previous administration in the U.S.