Cohere has released a new generation of its embedding models, the tools that turn documents into numbers so search systems can find the right passage. Embed 5 comes in two versions, Pro and Fast. Cohere’s changelog says both beat the earlier Embed 4 on financial filings, code, parsed PDFs, visually dense documents and multilingual search. The announcement includes no benchmark scores, so those gains are the company’s own unquantified claim.
The split is the practical part. Pro aims for retrieval accuracy and is meant for building the index. Fast aims for speed and suits live queries, including the rapid lookups agents make. The two share one embedding space, so a team can index with Pro and query with Fast, which is the pairing Cohere recommends. Either model accepts text, images or mixed pages, covers more than 100 languages, and reads up to 128,000 tokens at once. Vectors come in six sizes, from 256 to 2,048 dimensions.
Embed 5 is offered through Cohere’s own API, Microsoft Foundry and Amazon SageMaker, with Model Vault for single-tenant hosting. Teams paying for Embed 4 should test on their own documents before re-indexing, since no independent numbers exist yet.
Reported by Cohere’s documentation changelog on 30 September 2026.
