Llama 4

Llama 4 is Meta's open-weight model family built on a Mixture-of-Experts architecture, released in two downloadable variants: Scout, with 17 billion active parameters, 16 experts, and 109 billion total parameters, and Maverick, with 17 billion active parameters, 128 experts, and 400 billion total. Both models are natively multimodal, accepting text and image input, and Scout supports a context window of up to 10 million tokens, which suits long-document analysis and reasoning over large codebases. The family officially covers a dozen languages. Weights are distributed through Meta's site and Hugging Face under the Llama 4 Community License, which permits commercial use with restrictions for very large platforms. Developers typically run the models through inference stacks such as vLLM or Transformers, and they serve as open foundations for chat assistants, agents, and fine-tuned domain systems.