Meta Platforms Inc (NASDAQ:META, ETR:FB2A, SWX:FB) has released two new models from its Llama 4 suite, Scout and Maverick, which are open-source and designed for multimodal tasks, capable of processing text, images, audio, and video.
Llama 4 Scout runs on a single Nvidia H100 GPU using Int4 quantization, with 17 billion active parameters and 16 experts. It supports a 10-million-token context window and outperforms Gemini 3 and Mistral 3.1 on several benchmarks, according to a blog post from Meta.
Llama 4 Maverick also uses 17 billion active parameters but with 128 experts, targeting tasks like coding, reasoning, and multilingual applications.
Scout and Maverick are now available for download via Meta’s platforms and Hugging Face. The models are already being integrated into Meta’s product ecosystem, enhancing AI-driven features in WhatsApp, Messenger, and Instagram Direct.
A third model, Llama 4 Behemoth, is in training with 288 billion active parameters and is focused on STEM benchmarks.
“This is just the beginning for the Llama 4 collection. We believe that the most intelligent systems need to be capable of taking generalized actions, conversing naturally with humans, and working through challenging problems they haven’t seen before,” Meta wrote in the blog post.
“These models represent the best of Llama, offering multimodal intelligence at a compelling price while outperforming models of significantly larger sizes.”