Source: Cohere embeddings documentation

Cohere's guide to its Embed models (embed-v5.0-pro and embed-v3.0): input_type, Matryoshka dimensions, compressed embedding types and image embeddings.

Contents

Summary

As of November 2026 the page covers embed-v5.0-pro, a multilingual multimodal model, and the older embed-v3.0 family. Cohere makes the task explicit through input_type, offers Matryoshka dimensions, and can return several compressed representations in one call.

Key claims

  • input_type: search_query, search_document, classification, clustering (and image for v3.0); it tailors the embedding to its role.
  • embed-v5.0-pro supports output dimensions 256, 512, 768, 1024, 1536 and 2048.
  • embedding_types: float, int8, uint8, binary, ubinary; several can be requested at once. A 1024-dim binary vector is 128 bytes and may need unpacking for some vector stores.
  • embed-v5.0-pro can embed text and images together in one vector (e.g. slides, document screenshots); up to 96 images per request.

Feeds

Embeddings

Suggest changes →