Embed 5 Pro

Search that has to cover contracts, scanned reports and slide decks in many languages is the job Embed 5 Pro is built for. Cohere released it on 30 September 2026 as the bigger of two embedding models, intended for classification and for enterprise search with retrieval-augmented generation.

What it can embed

Cohere's announcement says the model can embed plain text, images, text and images fused together, and documents that have been parsed. More than 100 languages are covered. European languages get special attention, and coverage is broader for Japanese, Chinese, Korean, Arabic, Farsi, Hindi, Bengali, Telugu, Indonesian and Thai. Up to 128,000 tokens of context can be supplied.

Vector sizes and storage

Vectors come in six sizes, from 256 up to 2048 dimensions (2048, 1536, 1024, 768, 512 and 256), and as float, int8 or binary values. That lets a team trade storage and cost against accuracy to suit the size of its index.

Where to run it

It can be reached through Cohere's own API and Model Vault, through Microsoft Foundry and Amazon SageMaker, and on the North platform, and it can be deployed privately. Integrations are listed for the leading vector databases and search engines. Text and images are priced separately by the million tokens. Cohere also continues to list its earlier Embed v3.

Results Cohere reports

On retrieval tests for visual documents, financial material and parsed PDFs, Cohere reports leading results, which are its own figures. Every detail here comes from Cohere's announcement and documentation, and no independent report or measurement of Embed 5 was found.