Scout: 16 experts (109B total, 17B active), 10M tokens—breadth-optimized for massive knowledge bases. Maverick: 128 experts (400B total, 17B active), 1M tokens—depth-optimized for specialized reasoning.
Both activate ~17B parameters, making inference equally fast. But their design philosophies differ. Scout's 16 broad experts handle diverse token types flexibly—ideal for heterogeneous documents (contracts + emails + PDFs mixed). Maverick's 128 specialized experts target homogeneous, complex content—better for single domains (all code, all papers). With Zilliz Cloud, this matters: Scout adapts to varied retrieval (1000 mixed documents), Maverick specializes on focused retrieval (100 narrowly-scoped documents).
Context window is the differentiator: Scout's 10M tokens enable single-pass synthesis across thousands of documents. Maverick's 1M tokens require more selective retrieval (pre-filtering by Zilliz). Neither is universally better—choose by your problem. If Zilliz typically returns 500+ relevant documents per query, Scout. If Zilliz returns 50–100, Maverick. Both integrate identically with Zilliz Cloud; the choice is your retrieval volume and reasoning depth priority.
Related Resources
- Zilliz Cloud — Managed Vector Database — retrieve at your scale
- Retrieval-Augmented Generation (RAG) — model selection strategy
- Vector Embeddings — retrieval for both models