MongoDB drops Voyage 4 embedding models to solve the 'silent failure' of enterprise AI

Your AI agent is only as good as the data it can actually find. As agentic and RAG systems move from labs to production, retrieval quality is emerging as a silent failure point—one that MongoDB aims to fix with its newest Voyage 4 embeddings and reranking models.

Why MongoDB Voyage 4 embedding models are a game-changer

MongoDB has released four distinct versions of the Voyage 4 series to cover every enterprise use case. The flagship voyage-4-large is designed for maximum accuracy, while voyage-4-lite prioritizes low latency and cost-efficiency. For developers working in local or on-device environments, the voyage-4-nano marks MongoDB's first foray into open-weight models.

The performance claims aren't just marketing talk. On Hugging Face's RTEB benchmark, Voyage 4 clinched the top spot, outperforming rival models from Google and Cohere. Frank Liu, product manager at MongoDB, noted that getting embedding models right makes an application feel like it actually understands the user, rather than just delivering "random and shallow" search results.

Handling the complexity of multimodal data

Alongside the text models, MongoDB launched voyage-multimodal-3.5. This model is built to digest complex enterprise documents containing images, video, and text. It can extract semantic meaning from tables, charts, and figures—data that often gets lost in traditional RAG pipelines. It's a significant move for companies whose most valuable data isn't just plain text but hidden in slides and graphics.