Cohere Embed 4

Embed 4, released April 15, 2025, is Cohere's flagship embedding model and one of the first production multimodal embedders: it vectorizes text, images, and interleaved documents — PDFs with tables, slides with diagrams — into a single semantic space, no parsing pipeline required. The model IDs are embed-v4.0 (multimodal) and embed-english-v4.0 (text), with 128K context (roughly 200 pages), Matryoshka dimensions (256/512/1024/1536), and 100+ language coverage.

Cohere reported state of the art on text-to-text, text-to-image, and text-to-mixed-modality retrieval, and positioned it for agentic search over enterprise document mountains. It is available on the Cohere platform, Azure AI Foundry, Amazon SageMaker, and Bedrock, with VPC/on-premise private deployment for regulated industries.

Honest caveats: API-only with no open weights, benchmark claims are Cohere's, and multimodal embedding quality on niche document types deserves your own evaluation. For enterprise RAG over messy real-world documents, it is the reference multimodal embedder.