Meta AI gets two new models as Meta releases Llama 4

Meta has announced the merchandise of Llama 4, its newest postulation of AI models that present powerfulness Meta AI on the web and successful WhatsApp, Messenger, and Instagram Direct. The 2 models, besides disposable to download from Meta oregon Hugging Face now, are Llama 4 Scout, a tiny exemplary susceptible of “fitting successful a azygous Nvidia H100 GPU,” and Llama 4 Maverick, which is much akin to GPT-4o and Gemini 2.0 Flash. And the institution says it’s successful the process of grooming Llama 4 Behemoth, which Meta CEO Mark Zuckerberg says connected Instagram is “already the highest performing basal exemplary successful the world.”

According to Meta, Scout has a 10-million-token discourse model — the moving representation of an AI exemplary — and beats Google’s Gemma 3 and Gemini 2.0 Flash-Lite models, arsenic good arsenic the open-source Mistral 3.1, “across a wide scope of wide reported benchmarks,” portion inactive “fitting successful a azygous Nvidia H100 GPU.” It makes akin claims astir its larger Maverick model’s show versus OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, and says its results are comparable to DeepSeek-V3 successful coding and reasoning tasks utilizing “less than fractional the progressive parameters,” oregon the variables that usher AI models’ behavior.

Meanwhile, Llama 4 Behemoth has 288 cardinal progressive parameters with 2 trillion parameters successful total. The institution again says Behemoth tin outperform its competitors, successful this lawsuit GPT-4.5 and Claude Sonnet 3.7, “on respective STEM benchmarks.”

For Llama 4, Meta says it switched to a “mixture of experts” (MoE) architecture, an attack that conserves resources by utilizing lone the parts of a exemplary that are needed for a fixed task. The institution plans to sermon aboriginal plans for AI models and products astatine LlamaCon, which is taking spot connected April 29th.

As with its past models, Meta calls the Llama 4 postulation “open-source,” though it has been criticized for its licenses’ less-than-open requirements. For instance, the Llama 4 license requires commercialized entities with much than 700 cardinal monthly progressive users to petition a licence from Meta earlier utilizing its models, which the Open Source Initiative wrote successful 2023 takes it “out of the class of ‘Open Source.’”