Llama-4-Scout-17B-16E-Instruct-FP8

Scout is designed as a multimodal instruction model for long-running conversations and document-oriented work. Its mixture-of-experts design activates 17 billion parameters per token, combining broad model capacity with selective computation; the sources identify native multimodal handling and a nominal 10-million-token context window, making it suitable for workloads that need to process unusually large sequences in one interaction.

The practical appeal is flexible deployment rather than a single specialized strength. The model supports function calling, JSON-formatted responses, and streaming in the documented gateway setup, while Baseten’s preset illustrates vLLM deployment on four H100 GPUs with a 128,000-token serving configuration. The distinction between the advertised context capacity and a particular serving configuration is important when planning hardware, latency, and production throughput.