The Voyage 4 model family: shared embedding space with MoE architecture

The Voyage 4 series introduces a pioneering set of text embedding models that feature industry-first shared embedding spaces, including models like voyage-4-large, voyage-4, voyage-4-lite, and voyage-4-nano. These models allow for interchangeable use of embeddings, providing flexibility for users to optimize for accuracy, latency, and cost by combining different models for query and document embedding. The flagship model, voyage-4-large, employs a mixture-of-experts architecture, achieving state-of-the-art retrieval accuracy with costs significantly lower than dense models. The series supports dimensional embeddings and employs Matryoshka learning and quantization to maintain high retrieval accuracy while reducing database costs. Evaluations using the Retrieval Embedding Benchmark demonstrate that voyage-4-large outperforms other models like Gemini Embedding 001 and Cohere Embed v4, with asymmetric retrieval enhancing quality when smaller models retrieve documents embedded by voyage-4-large. The models are accessible through the Voyage API and MongoDB Atlas, with voyage-4-nano available on Hugging Face for local development.