Llama 4 Scout

Llama 4 Scout is a language model Meta released on 5 April 2025 as the smaller of the two published Llama 4 models. It uses a mixture-of-experts architecture with 16 experts, roughly 109 billion parameters in total and 17 billion active per token, takes text and image input natively, and is documented with a context window of up to 10 million tokens. Its efficiency comes from sparse mixture-of-experts routing and from codistillation with Llama 4 Behemoth: only a fraction of the weights is active per token, so compute per token tracks the 17 billion active parameters rather than the full 109 billion. Meta states that Scout fits on a single NVIDIA H100 GPU with on-the-fly Int4 quantization, which sets its deployment profile: one high-end accelerator, not edge hardware. Weights are published under the Llama 4 Community License, so the model can be self-hosted in a private cloud or used through hosting providers; at roughly 109 billion total parameters it is not a model for phones or embedded devices. Common uses are assistant and support systems, document processing over long inputs, content generation, and code assistance, where the long context window and single-accelerator footprint matter more than maximum model size. The 10-million-token figure is a documented maximum rather than a promise of quality at that length; serving providers commonly expose a far smaller context window, so the usable limit depends on the deployment.

Related terms

• Code Llama
• LangChain Llama 3
• Llama 3
• Llama 4 Behemoth
• Llama 4 Maverick

Related services: Agentic AI consulting.

← Back to the AI Glossary