Voyage AI Models

The new flagship MoE model. Establishes a new retrieval accuracy frontier while maintaining lower serving costs than comparable dense models.

Approaches the retrieval accuracy of the previous flagship model voyage-3-large while maintaining the high efficiency of a mid-sized model. All voyage-4 models use a shared embedding space.

Enables high-quality embeddings at significantly reduced computational cost, approaching the retrieval accuracy of voyage-3.5.

A lightweight, open-weights embedding model optimized to run on local developer machines.

The single model for text, images, and video. Explicitly trained on video frames to unlock semantic search for video archives, documents, and screenshots without sacrificing text accuracy.

Contextualized chunk embedding model that produces vectors for chunks that capture the full document context without manual metadata and context augmentation.

rerank-2.5 is a general-purpose reranker optimized for retrieval quality with multilingual support.

rerank-2.5-lite is a general-purpose reranker optimized for latency & quality with multilingual support.