Meta Llama 4 Scout

Llama 4 Scout: The 17B-parameter, multimodal MoE model with a massive 10 million-token context, built for deep document and image analysis.

Meta Llama 4 Scout is the high-efficiency, natively multimodal AI model engineered for heavy-duty retrieval tasks. It leverages a Mixture-of-Experts (MoE) architecture, activating 17 billion parameters from 16 experts for each forward pass. The key differentiator is its industry-leading 10 million-token context window: this capacity allows developers to process and synthesize insights from entire knowledge bases or years of documentation in a single prompt. Optimized for performance and efficiency, Scout is designed to run effectively on a single Nvidia H100 GPU, offering best-in-class results against rivals like Google's Gemma 3 and Mistral 3.1 in benchmark tests. This model is your go-to solution for advanced Retrieval-Augmented Generation (RAG) and complex text/image understanding.

https://llama.com