Embed v4

Cohere's fourth-generation multimodal embedding model processes text and images together in a single shared vector space, eliminating the need for separate specialized models across content types. The architecture supports Matryoshka representation learning, which allows nested truncation to smaller vector dimensions with controlled quality loss—a capability designed for flexible cost-versus-accuracy tradeoffs at query time. It achieved a 65.2 MTEB benchmark score, outpacing comparable models like OpenAI's text-embedding-3-large at 64.6, which positions it as a strong choice for semantic search and classification tasks over heterogeneous documents.

The model natively embeds visual content—screenshots of PDFs, slides, figures, and tables—alongside text without requiring preprocessing to convert visual elements into text first. That removes a common bottleneck in document-heavy retrieval pipelines where mixed-format assets dominate: technical documentation with diagrams, investor presentations with charts, research reports with figures. Its flexible output dimensions support tiered retrieval architectures, using compressed vectors for fast candidate filtering and full-resolution embeddings for precision re-ranking. With open weights available, organizations can fine-tune the model for domain-specific retrieval scenarios, making it particularly suited for building agentic AI applications over multimodal knowledge bases.