Cohere on April 15, 2025 published a changelog announcing Embed Multimodal v4 (Embed v4), a release aimed at improving search and retrieval for multimodal enterprise workflows rather than serving as a general — purpose chat model. The update matters because it combines configurable multi — dimensional embeddings with a very long 128k input context, which directly targets use cases that require indexing and retrieving mixed — content artifacts at scale.
Cohere reports that Embed v4 achieves state — of-the-art performance in three retrieval tasks: text-to-text retrieval, text-to-image retrieval, and text-to-mixed modality retrieval. The company highlights mixed — modality retrieval with concrete examples such as PDFs that contain both text and images, emphasizing the model’s suitability for documents and artifacts where content types are interleaved. The model is available immediately on the Cohere Platform and via cloud integrations on AWS SageMaker and Azure AI Foundry. Cohere points readers to a dedicated blog post and the changelog entry for implementation details and usage guidance; the changelog entry announcing the release is dated April 15, 2025.
For builders, Embed v4 creates practical operational choices. Matryoshka Embeddings let teams select among multiple vector sizes to trade off storage, latency and retrieval fidelity depending on application needs. Unified multimodal outputs simplify pipelines that would otherwise require separate image and text encoders and a post-processing step to combine vectors. The 128k context expands options for retrieval and indexing strategies over long documents and combined — content artifacts like PDFs. Together, these changes aim to streamline mixed — content search workflows for enterprises that need scalable retrieval over long, multimodal documents. For implementation specifics and examples, Cohere’s changelog and the linked blog post provide usage guidance and technical details.
Sources
Replies (0)
No replies in this topic yet.
