Embed v4

Embed v4 is Cohere's most performant search model to date, built as an enterprise-focused embedding engine that transforms text, images, and mixed content into vector representations for semantic search and classification. Its unified embeddings approach handles multimodal input in a single payload, meaning images and text can be embedded together rather than processed separately. The model introduces Matryoshka Embeddings with variable output dimensions, allowing developers to select smaller vectors when full precision is unnecessary, trading accuracy for speed and storage efficiency. It achieves state-of-the-art results across text-to-text retrieval, text-to-image retrieval, and text-to-mixed-modality retrieval from complex documents like PDFs.

Part of a family of eight embedding models released between 2023 and 2025, Embed v4 represents the culmination of Cohere's ongoing work in this space. The model's open-weight availability makes it attractive for teams that want to deploy custom retrieval pipelines without vendor lock-in. Its exceptionally large context window enables a fundamentally different approach to retrieval-augmented generation, where entire technical specifications, annual reports, legal contracts, or research papers can be embedded as complete units rather than fragmented chunks. This whole-document capability preserves semantic coherence that traditional chunking strategies lose, and it has driven adoption across platforms including Azure AI Foundry and AWS SageMaker for enterprise-scale search deployments.