Scout supports 10M tokens for massive knowledge base retrieval; Maverick supports 1M tokens for dense reasoning on focused domains.
Maverick's 128 experts (400B total, 17B active) are specialized for depth—each expert might focus on reasoning, language, math, or domain patterns. Scout's 16 experts (109B total, 17B active) are generalists for breadth. In RAG terms: Scout absorbs thousands of documents and synthesizes, Maverick deeply reasons about hundreds of documents. For teams using Zilliz Cloud, this matters at scale. Zilliz manages vector indexing and retrieval; you choose your reasoning model. Scout is ideal when your knowledge base is massive (1M+ documents, retrieve 500+ per query) and you need comprehensive answers. Maverick shines for specialized domains (financial analysis, medical research, code understanding) where expert routing matters and context is pre-filtered.
Both activate only 17B parameters, so inference cost and latency are identical at the model level. The difference is retrieval strategy. With Scout, you can retrieve aggressively (top-1000 vectors) without fear—Scout processes them all. With Maverick, precise retrieval (top-50 vectors) is critical because context is bounded. Choose Scout for breadth, Maverick for depth.
Related Resources
- Zilliz Cloud — Managed Vector Database — scale retrieval a search engine providerally
- Retrieval-Augmented Generation (RAG) — RAG architecture
- Vector Embeddings — embedding foundations for retrieval