Llama 4 Scout 17B

Llama 4 Scout 17B is built on a mixture-of-experts architecture that activates 17 billion parameters from a much larger 109 billion parameter model, routing through 16 specialized experts to maximize efficiency. This design enables high-throughput inference while maintaining strong performance across language and vision tasks. The model incorporates early fusion, allowing text and image tokens to be processed together from the start, which creates more seamless multimodal reasoning. Scout is designed for grounded visual comprehension at scale, making it particularly suited for applications that require detailed language reasoning based on visual context.

The model was instruction-tuned to support assistant-style interaction across 12 supported languages, spanning multilingual chat, captioning, and image understanding. Training involved a mix of publicly available data, licensed content, and information from Meta's products, totaling approximately 40 trillion tokens with a knowledge cutoff in August 2024. Scout excels at multi-document summarization, parsing extensive user activity for personalized tasks, and reasoning over vast codebases. Built for both local and commercial deployment, it is available as an open-weight model under the Llama 4 Community License, enabling developers to self-host or integrate it into enterprise workflows with flexibility.