Scout: 16 experts (109B total, 17B active) optimized for breadth and long contexts. Maverick: 128 experts (400B total, 17B active) optimized for expert specialization and depth.
Both activate ~17B parameters per token, but route differently. Scout’s 16 experts are broad—each handles diverse reasoning types. Maverick’s 128 experts are specialized—individual experts might focus on math, language structure, commonsense, etc. For Milvus users, this affects how retrieved content is processed. Scout’s broad experts handle varied document types (contracts, emails, PDFs, code) flexibly. Maverick’s specialized experts excel at homogeneous, complex content (all medical papers, all code repositories) where expert specialization pays off.
Mixture-of-experts design in both means you get dense model quality with sparse inference cost. The gating network learns which experts to activate for each token type. Scout’s 10M context window is its differentiator for massive retrieval; Maverick’s 1M window paired with 128 experts excels at deep reasoning. In Milvus deployment: Scout handles breadth (retrieve 1000 documents, process all); Maverick handles depth (retrieve 50 documents, reason thoroughly). Both are equally fast at inference because activation is sparse—choose by your RAG problem shape, not by speed.
Related Resources
- Milvus Quickstart — evaluate both architectures
- Enhance RAG Performance — architecture selection strategy
- RAG with LlamaIndex — integration for both models