Llama 4 Scout
Overview
Llama 4 Scout is an open-weight, efficient AI model developed by Meta AI, released on April 5, 2025. It utilizes a Mixture of Experts (MoE) architecture, activating 16 experts per forward pass, and supports a context window of 10 million tokens, enabling processing of extensive documents and extended conversations. The model is natively multimodal, handling both text and image inputs, and is instruction-tuned for tasks such as multilingual chat, captioning, and image understanding. Llama 4 Scout is available under the Llama 4 Community License, allowing for self-hosting and integration into various applications. Its design emphasizes efficiency, capable of running on a single NVIDIA H100 GPU with heavy quantization, making it suitable for cost-sensitive deployments. In terms of performance, Llama 4 Scout demonstrates competitive capabilities across several benchmarks. It achieves an Arena Elo score of 1,322, indicating strong performance in conversational tasks. The model's MMLU Pro score is 80.4%, reflecting its proficiency in a range of reasoning and knowledge tasks. In coding tasks, as measured by HumanEval, it attains a score of 76.3%, showcasing its coding capabilities. Additionally, Llama 4 Scout has a context window of 10 million tokens, allowing it to process extensive documents and extended conversations. These attributes position Llama 4 Scout as a versatile and efficient model suitable for applications requiring long-context processing and multimodal understanding, while maintaining competitive performance across various tasks.