Technical Specifications

System Requirements

VRAM requirements for different quantization methods and context sizes

Llama 4 Scout

Llama 4 Scout is a key offering within Meta's Llama 4 family of models, released on April 5, 2025. It is designed to provide robust artificial intelligence capabilities for researchers and organizations while operating within practical hardware constraints. As a general-purpose model, Llama 4 Scout exhibits native multimodality, proficiently processing both text and image inputs. Its applications encompass a wide array of tasks, including complex conversational interactions, detailed image analysis, and advanced code generation. The model's design focuses on enabling efficient execution of these tasks across diverse computational environments.

Architecturally, Llama 4 Scout employs a Mixture-of-Experts (MoE) configuration, incorporating 109 billion total parameters, with 17 billion active parameters engaged per token across 16 experts. A significant innovation in its design is an industry-leading context window, supporting up to 10 million tokens, which represents a substantial increase over prior iterations. The model integrates an early fusion approach for its native multimodality, which unifies text and vision tokens within its foundational structure. Optimized for efficient deployment, Llama 4 Scout can run on a single NVIDIA H100 GPU when leveraging Int4 quantization. Furthermore, its architecture incorporates interleaved attention layers, specifically iRoPE, to enhance generalization capabilities across extended sequences.

Llama 4 Scout is well-suited for applications demanding the processing and analysis of extensive information volumes. Its primary use cases include multi-document summarization, detailed analysis of user activity for personalization, and reasoning over substantial codebases. The model demonstrates strong performance in tasks requiring document question-answering, precise information retrieval, and reliable source attribution, making it particularly valuable for professional document analysis. Its design for efficiency on a single GPU facilitates accessibility for organizations with varying computing infrastructure. The model also supports multilingual tasks, having been trained on data from 200 languages, with fine-tuning capabilities for 12 specific languages.