# Cohere’s Embed 4 put text, images and mixed documents in one retrieval representation Cohere released Embed 4 on April 15, 2025, combining mixed text-image embeddings, 128K context and selectable embedding dimensions for retrieval. Canonical: https://rangoon.ai/insights/archive-cohere-embed-4-april-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-23 Updated: 2026-10-23 Event date: 2025-04-15 Tags: [Cohere](https://rangoon.ai/insights/tags/cohere/), [Embed 4](https://rangoon.ai/insights/tags/embed-4/), [multimodal retrieval](https://rangoon.ai/insights/tags/multimodal-retrieval/), [vector search](https://rangoon.ai/insights/tags/vector-search/) Source note: Historical reporting: Cohere released Embed 4 on April 15, 2025. The launch notes name mixed image-and-text embeddings, a 128K context length, dimensions of 256, 512, 1024, and 1536, and availability through Cohere, AWS SageMaker, and Azure AI Foundry. The current Embed documentation provides later interface context; no vendor retrieval-quality claim is treated as an independent benchmark. Image: Multimodal embedding and retrieval design boundary. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Embed 4 created one embedding representation from text, images, or mixed inputs, according to Cohere. - Its selectable dimensions expose a storage-and-compute tradeoff that still needs corpus-specific retrieval testing. - Platform availability is separate from an application’s indexing, metadata, access-control and ranking design. ## The April launch joined mixed inputs with a configurable vector size [source 1](https://docs.cohere.com/changelog/embed-multimodal-v4) Cohere announced Embed 4 on April 15, 2025. The release said it could produce unified embeddings from text, images, or mixed input in one payload, support a 128K context length, and return Matryoshka embeddings at 256, 512, 1024, or 1536 dimensions. Cohere said the model was then available on its platform, AWS SageMaker, and Azure AI Foundry. An embedding is a numerical representation used to compare semantic similarity; it is not a generated answer or a retrieval system by itself. The launch explains what representation Cohere offered, but it does not show the quality of a customer index, the contents of a corpus, or the ranking behavior of a full search experience. ## One vector space can preserve more of a mixed document than text extraction alone [source 1](https://docs.cohere.com/changelog/embed-multimodal-v4) [source 2](https://docs.cohere.com/v2/docs/cohere-embed) Mixed documents often contain prose, figures, tables, scans, and layout cues that do not map cleanly to one text string. A unified embedding interface lets a retrieval design represent those inputs without first reducing every page to plain text. Current Cohere documentation still lists embed-v4.0 for text, image, and mixed text-image inputs, which is later maintained interface context. That representation choice moves a system question upstream rather than solving it automatically. A team still decides how to segment a document, attach source metadata, store access controls, generate query embeddings, and return a result with enough provenance for a user to inspect it. ## Selectable dimensions make index cost a design variable [source 1](https://docs.cohere.com/changelog/embed-multimodal-v4) The four announced vector lengths give an implementer a knob for index footprint and similarity-computation cost. Shorter vectors can reduce storage and some search work, while longer vectors retain more coordinates for a retrieval method to use. The release does not say that one length is universally better for every language, content type, query distribution, or vector index. Matryoshka support does not make a validation set optional. Retrieval quality depends on the documents, chunk boundaries, query intent, distance metric, approximate-nearest-neighbor settings, filters, and any later reranking step, so the viable dimension should be measured on the intended task. ## Model availability is not deployment or relevance proof [source 1](https://docs.cohere.com/changelog/embed-multimodal-v4) [source 2](https://docs.cohere.com/v2/docs/cohere-embed) Cohere’s April notes establish its stated cloud distribution channels for Embed 4. They do not establish that a model was enabled in every account, that a particular region or tenancy was ready, or that a customer had indexed its materials. They also do not turn Cohere’s product-performance language into an independent comparison. The event supports a focused historical claim: Embed 4 introduced the stated multimodal representation and dimension options. It does not prove a retrieval gain for a particular PDF collection, accuracy for a regulated workflow, or a complete answer-generation system. ## Sources 1. [Cohere release notes: Announcing Embed Multimodal v4](https://docs.cohere.com/changelog/embed-multimodal-v4) 2. [Cohere documentation: Cohere’s Embed Models](https://docs.cohere.com/v2/docs/cohere-embed) ## Related reading - [Jina’s v4 release combined text-image embeddings with dense and late-interaction outputs](https://rangoon.ai/insights/archive-jina-embeddings-v4-june-2025/) - [Google made Gemini Embedding stable while keeping vector size a retrieval choice](https://rangoon.ai/insights/archive-google-gemini-embedding-general-availability-july-2025/) - [Cohere released Rerank 4 with Pro and Fast variants for multilingual and JSON document ranking](https://rangoon.ai/insights/archive-cohere-rerank-4-december-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)