Llama 4 Scout 17B Instruct (NovitaAI)

Llama 4 Scout 17B Instruct is a member of the Llama family built around a sparse mixture-of-experts design that activates 17 billion parameters per forward pass while keeping the total parameter count at 109 billion spread across 16 expert pathways. This routing strategy lets the model direct work to specialized subnetworks instead of engaging the full architecture on every query, which is described as the reason it can deliver strong performance while remaining manageable enough for practical deployment scenarios. The Instruct tuning shapes the model for assistant-style interaction, making it a natural fit for conversational assistants, retrieval-augmented reasoning, code-oriented helpers, and other chat-driven applications where responsive natural language generation matters.

A defining feature of the architecture is its early-fusion approach to multimodality, which blends text and image tokens into a unified representation from the start of processing rather than treating vision as a downstream add-on. This unified tokenization supports seamless multimodal reasoning across documents, diagrams, and multi-image inputs, and it pairs with broad multilingual coverage that spans twelve languages for both understanding and generation across text and code. The combination of sparse activation, early fusion, and instruction tuning positions the model as a versatile option for users who want an open-weights Llama 4 variant capable of handling rich, mixed-media assistant workloads without requiring the largest dense model in the family.