Cohere released Embed 5 on [date], saying it delivers stronger retrieval across complex enterprise data while giving teams more control over latency, cost, and deployment. It is the company's first major update to its embeddings models since the release of Embed 42.
Cohere reported an average score of 85.8 on ViDoRe V3, measured across financial documents, parsed PDFs, and image retrieval. That compares with 83.7 for Voyage 4 Large.
Embed 5 is built on Matryoshka representations and lower-precision outputs, targeting enterprise data with multimodal inputs, 100+ languages, and a 128K-token context window. Availability begins today, initially for developers and enterprise users.
"Embed 5 delivers our strongest retrieval performance to date," said [name and title exactly as the source gives them]. [One sentence paraphrasing what that person went on to say, from the source only].
The announcement follows the release of Embed 42. [One sentence on how the source itself frames the significance, adding no judgement of your own].
Cohere did not say how the models will perform on untested benchmarks, and [the open question or limitation the source raises]. [One sentence on what the source says happens next].
Source: cohere
