Meta has announced the release of Llama 4, its latest assortment of AI fashions that now energy Meta AI on the web and in WhatsApp, Messenger, and Instagram Direct. The 2 fashions, additionally accessible to obtain from Meta or Hugging Face now, are Llama 4 Scout, a small mannequin able to “becoming in a single Nvidia H100 GPU,” and Llama 4 Maverick, which is extra akin to GPT-4o and Gemini 2.0 Flash. And the corporate says it’s within the course of of coaching Llama 4 Behemoth, which Meta CEO Mark Zuckerberg says on Instagram is “already the best performing base mannequin on the earth.”
In line with Meta, Scout has a 10-million-token context window — the working reminiscence of an AI mannequin — and beats Google’s Gemma 3 and Gemini 2.0 Flash-Lite fashions, in addition to the open-source Mistral 3.1, “throughout a broad vary of extensively reported benchmarks,” whereas nonetheless “becoming in a single Nvidia H100 GPU.” It makes comparable claims about its bigger Maverick mannequin’s efficiency versus OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, and says its outcomes are corresponding to DeepSeek-V3 in coding and reasoning duties utilizing “lower than half the energetic parameters,” or the variables that information AI fashions’ conduct.
In the meantime, Llama 4 Behemoth has 288 billion energetic parameters with 2 trillion parameters in complete. The corporate once more says Behemoth can outperform its opponents, on this case GPT-4.5 and Claude Sonnet 3.7, “on a number of STEM benchmarks.”
For Llama 4, Meta says it switched to a “combination of specialists” (MoE) structure, an strategy that conserves assets by utilizing solely the components of a mannequin which are wanted for a given activity. The corporate plans to debate future plans for AI fashions and merchandise at LlamaCon, which is taking place on April 29th.
As with its previous fashions, Meta calls the Llama 4 assortment “open-source,” though it has been criticized for its licenses’ less-than-open necessities. As an example, the Llama 4 license requires industrial entities with greater than 700 million month-to-month energetic customers to request a license from Meta earlier than utilizing its fashions, which the Open Supply Initiative wrote in 2023 takes it “out of the class of ‘Open Supply.’”