Cohere launches Embed 5 to search text and images together

Cohere introduced the Embed 5 family of embedding models on September 30, 2026. The family has two models: Embed 5 Pro, which puts quality first, and Embed 5 Fast, built for speed and high volume. Both turn text, images and inputs that combine the two into the same numerical representation, so scanned documents, tables and charts can be found in the same search as text.

What an embedding model does

An embedding model converts a piece of text or an image into a sequence of numbers (a vector) that carries its meaning. Search systems and RAG (retrieval-augmented generation) applications use these vectors to find the documents closest in meaning to a query. Since much of a company archive consists of scanned PDFs, slides, tables and charts, an embedding model that understands images directly can reduce the need to convert documents to text first.

Technical details

According to Cohere, both models handle up to 128K tokens of context and support more than 100 languages. Output dimensions can be chosen from 2048 down to 256. Thanks to Matryoshka representation learning, the smaller sizes come from the first part of the full vector, which helps cut storage and search costs. Vectors are available as float, int8 and binary.

Pro and Fast share the same embedding space: documents indexed with Pro can be queried with Fast without re-indexing. Cohere recommends Pro for offline indexing and quality-critical retrieval, and Fast for interactive search, high-volume RAG and agentic retrieval. In the results Cohere shared, Embed 5 Pro comes out ahead of rival models such as Voyage and Gemini Embedding on scanned document and page-image retrieval, financial datasets and parsed PDF tests.

Access and deployment options

The models are available through the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and North, Cohere's enterprise AI platform. Cohere says both models can be served with vLLM and run as private deployments in a company's own virtual private cloud (VPC) or on premises. Embed 5 has not been published on Hugging Face for anyone to download, so it does not fall into the open model category. Current pricing is on Cohere's website.

An option for companies with PDF and chart archives

The main audience for Embed 5 is companies building search or RAG in fields such as finance, law, insurance and healthcare, where much of the information sits in scanned documents, tables and charts. Support for more than 100 languages also matters for organizations working with documents in Turkish or Arabic. For companies in regulated sectors that do not want their data to leave, the option to run on their own infrastructure is an added point.

Part of the comparison was done with RCP-nDCG@10, Cohere's own evaluation method introduced the same day. It scores retrieved documents against relevance criteria specific to each query. Results may differ on real company archives, so it is worth testing the models on your own documents and following evaluations from independent researchers.

Sources: Cohere Blog: Embed 5.