Meta released the first two Llama4 models: Llama4 Scout and Llama4 Maverick. Maverick, the “workhorse” among the two, excels in image and text recognition for “general assistant use cases and chat,” according to the company. img alt=””A text slide describing three models from the Llama 4 family: Llama 4 Behemoth, Llama 4 Maverick and Llama 4 Scout” ” height=””540″ ” src=””https://s.yimg.com/ny/api/res/1.2/OUDoX9CXbSgsMAjqidWXOg–/YXBwaWQ9aGlnaGxhbmRlcjt3PTk2MDtoPTU0MA–/https://s.yimg.com/os/creatr-uploaded-images/2025-04/215de090-1266-11f0-aa7d-5ca8559aedef” ” width=””960″”/>
img alt=””A text slide describing three models from the Llama 4 family: Llama 4 Behemoth, Llama 4 Maverick and Llama 4 Scout””height=””540″”src=””https://s.yimg.com/ny/api/res/1.2/OUDoX9CXbSgsMAjqidWXOg–/YXBwaWQ9aGlnaGxhbmRlcjt3PTk2MDtoPTU0MA–/https://s.yimg.com/os/creatr-uploaded-images/2025-04/215de090-1266-11f0-aa7d-5ca8559aedef””width=””960″”,/>
Scout has 17 billion active parameters with 16 experts and
(
Meta claims that Scout has 17 billion parameters and 16 experts. Maverick, on the other hand, has 17 billion parameters and 128 experts. According to Zuckerberg, it’s “extremely fast, natively Multimodal, has an industry-leading, nearly infinite 10,000,000 token context length and is designed to run using a single GPU.” The company claims that it is superior to competitors such as GPT-4o, Gemini 2.0, and DeepSeek v3.1 in coding, reasoning and multilingual benchmarks.
According to the company, Zuckerberg has already called the upcoming Behemoth base model, which is currently training, “the world’s highest performing model,” with 288 trillion active parameters. It’s not here yet, but we’ll likely hear a lot about it and the Reasoning model in the near future. Meta’s LlamaCon AI developer conference is just a few short weeks away.
