voyage-4-large

Voyage 4 Large represents a notable architectural shift for text embedding models, being the first production-grade embedding system to adopt a mixture-of-experts design. This MoE approach allows the model to deliver state-of-the-art retrieval accuracy while keeping serving costs roughly 40% lower than comparable dense architectures. Beyond raw efficiency, the model introduces an industry-first shared embedding space that spans the entire Voyage 4 family—meaning developers can index documents with the flagship model and query with any sibling model without re-encoding, enabling flexible quality-latency-cost tradeoffs tailored to each use case.

The model incorporates Matryoshka Representation Learning alongside quantization-aware training, supporting output dimensions from 2048 down to 256 and multiple data types including float32, int8, uint8, binary, and ubinary formats. This flexibility lets teams dial in performance for specific deployment constraints without sacrificing meaningful retrieval quality. Built for both traditional RAG pipelines and the emerging class of context-engineered agents with long-term memory, Voyage 4 Large balances frontier retrieval performance with practical serving economics for high-volume production workloads.