Llama 4 Scout 17B Instruct

Llama 4 Scout 17B Instruct is Meta's multimodal entry in the Llama 4 family, built on a mixture-of-experts architecture that activates 17 billion parameters per inference from a much larger pool of 109 billion total parameters spread across 16 expert modules. This MoE design lets the model keep per-request compute relatively modest while still drawing on broad learned capacity, making it well suited for assistant-style interactions that mix text reasoning with image understanding. As an open-weights release, Scout gives developers direct access to the model weights for self-hosting and fine-tuning, which is a meaningful fit for teams that want to customize a multimodal model without paying for closed-API access.

Practically, Scout is positioned for general multimodal workloads where a single model has to handle both text and image inputs and produce coherent text responses, including document and OCR-style vision tasks alongside conversational reasoning. The MoE structure means it can scale its effective knowledge base while controlling inference cost, which tends to suit production assistants that need broad capability coverage rather than narrow single-task specialization. Its open-weight availability also makes it a practical base for domain adaptation, retrieval-augmented setups, and on-premise deployments where data control matters, fitting well into developer pipelines that already use Llama-family tooling.