Voyage Launches Voyage 4 Series with Industry-First Shared Embedding Spaces and Mixture-of-Experts Architecture
Key Takeaways
- ▸Shared embedding space is an industry-first: all Voyage 4 models produce compatible embeddings, enabling flexible model mixing for queries and documents
- ▸Mixture-of-experts architecture in voyage-4-large achieves state-of-the-art retrieval accuracy with 40% lower serving costs than comparable dense models
- ▸Asymmetric retrieval strategy allows cost-efficient production deployments by using larger models for document embeddings (one-time cost) and smaller models for query embeddings (ongoing cost)
Summary
Voyage has announced the Voyage 4 model family, a new generation of text embedding models featuring an industry-first capability: shared embedding spaces that allow embeddings from different models to be used interchangeably. The series includes four models—voyage-4-large, voyage-4, voyage-4-lite, and the open-weighted voyage-4-nano—each optimized for different accuracy, latency, and cost requirements.
The flagship voyage-4-large model introduces a mixture-of-experts (MoE) architecture, marking the first production-grade embedding model to utilize MoE design. This approach delivers state-of-the-art retrieval accuracy while maintaining serving costs 40% lower than comparable dense models. The shared embedding space enables asymmetric retrieval strategies, where larger models vectorize document corpora once and smaller models handle continuous query embeddings, significantly reducing per-query costs.
Voyage is also releasing voyage-4-nano as an open-weight model on Hugging Face under the Apache 2.0 license, providing developers with a freely available option for local development and prototyping. The tiered model lineup allows production teams to independently optimize document embeddings for accuracy and query embeddings for latency and cost, with seamless upgrades between models without re-vectorizing data.
- Open-weight voyage-4-nano model freely available on Hugging Face accelerates adoption for local development and prototyping
Editorial Opinion
The Voyage 4 series represents a meaningful step forward in embedding model architecture and deployment flexibility. The shared embedding space is genuinely novel—it solves a real production problem by decoupling document and query embedding strategies while maintaining semantic compatibility. The MoE approach achieving state-of-the-art accuracy at lower cost challenges the assumption that frontier performance requires proportional infrastructure spending, and open-sourcing the nano model signals confidence while lowering barriers to adoption.