Meta has recently introduced a fresh set of AI models known as Llama 4. This new release includes four models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models underwent training on large volumes of unlabeled text, images, and videos to enhance their comprehensive visual perception, as stated by Meta.

The latest developments in open models from the Chinese AI lab, DeepSeek, which demonstrated comparable or better performance to Meta's previous leading Llama models, reportedly accelerated the progress of the Llama project. This prompted Meta to swiftly analyze how DeepSeek managed to reduce the operational expenses associated with models like R1 and V3.

Scout and Maverick models are readily accessible on Llama.com and through Meta's partners, like the AI developer platform Hugging Face, while Behemoth is still undergoing training. Llama 4 integration has been updated for Meta AI, the AI-driven assistant utilized in apps such as WhatsApp, Messenger, and Instagram, across 40 countries. Presently, the multimodal capabilities are exclusively available in the U.S. in English.

There have been concerns raised by some developers regarding the Llama 4 licensing terms. Notably, individuals and businesses based in the EU are restricted from using or distributing these models due to regulatory obligations concerning AI and data privacy laws within the region. Moreover, companies with over 700 million monthly active users must seek a special license from Meta, subject to approval or denial at Meta's discretion.

Meta expressed in a blog post that the Llama 4 models signify the commencement of a new chapter for the Llama ecosystem, hinting at more to come from the Llama 4 series.

Regarding technology, Meta indicates that Llama 4 represents the first set of models to leverage a mixture of experts (MoE) architecture, which is more computationally effective for training and responding to queries. This architecture divides data processing tasks into smaller, specialized "expert" models to enhance efficiency.

Maverick, as an example, comprises 400 billion total parameters, with only 17 billion active parameters distributed among 128 "experts." On the other hand, Scout features 17 billion active parameters, 16 experts, and a total of 109 billion parameters.

Internal testing conducted by Meta suggests that Maverick, tailored for general assistant and chat functions such as creative writing, outperforms models like OpenAI's GPT-4o and Google's Gemini 2.0 in coding, reasoning, multilingual tasks, long-context understanding, and image analysis assessments. However, Maverick falls short when compared to more advanced models like Google's Gemini 2.5 Pro, Anthropic's Claude 3.7 Sonnet, and OpenAI's GPT-4.5.