Cohere Embed iconCohere Embed

Cohere's multimodal embedding API: Embed 5 Pro and Fast with 128K context, text and images, 100+ languages, int8 and binary output

128K Context (Embed 4 and 5)
100+ Languages
2,048 Max dimensions (Embed 5)

Overview

Cohere Embed is the embedding model line from Cohere, built for enterprise search and RAG. The newest generation, Embed 5, comes in two sizes that share one vector space: embed-v5.0-pro (about 2.4B parameters) for indexing and embed-v5.0-fast (about 1B) for queries, so you can index with Pro and query with Fast. Both take up to 128K tokens of text, images or mixed text and image content such as PDF pages, cover 100+ languages, and output 256 to 2,048 dimensions as float, int8 or binary vectors. embed-v4.0 (launched April 2025, 128K, up to 1,536 dimensions) and the older v3 models (512 tokens) are still offered. You can run it on Cohere's API, AWS (Bedrock, SageMaker), Azure AI Foundry, or a private VPC or on-premises install. Also from Cohere: Cohere (Command models and platform), Cohere North (enterprise AI workspace) and Cohere Rerank (reranking models).

The Verdict

Who Should Use Cohere Embed?

Best For

  • Enterprises that need private VPC or on-premises embeddings
  • Multilingual search across 100+ languages
  • Indexing long documents and scanned PDFs as one vector
  • Teams already on AWS Bedrock or Azure AI Foundry

Not Ideal For

  • Buyers who want a public per-token price before signing up
  • Self-hosting on your own GPUs without a Cohere agreement
  • Hobby projects: trial keys cannot be used in production

What's Great

  • 128K context, far beyond OpenAI's 8,192 tokens
  • Text, images and mixed pages in one model
  • int8 and binary output; Cohere claims up to 96% storage compression
  • Pro and Fast share a vector space, so you can index with Pro and query with the smaller Fast
  • Strongest on non-English tickets in a 2026 hands-on test

Watch Out For

  • Cohere's pricing page shows no Embed price; you see rates only after you get a production key
  • embed-v4.0 scored Recall@10 0.85 vs 0.89 for voyage-4-large on an English corpus
  • One AWS Marketplace reviewer found v4 distance scores less useful than OpenAI's for RAG
  • Older models retire: Embed v2.0 was retired on 2026-04-04

Pricing

View all features & details

Current Models

  • embed-v5.0-pro and embed-v5.0-fast: 128K, 256 to 2,048 dims
  • embed-v4.0: 128K, 256 to 1,536 dims
  • embed-english-v3.0 and embed-multilingual-v3.0: 512 tokens, 1,024 dims
  • Light v3 variants: 512 tokens, 384 dims

Where It Runs

  • Cohere API
  • Amazon Bedrock (Embed 4 since October 2025) and SageMaker
  • Azure AI Foundry
  • Private VPC and on-premises

Related Product

  • Cohere Rerank 5 (Pro and Fast) is sold separately and billed per search (one query with up to 100 documents)

Real-World Usage

Independent Test (2026)

  • English RAG corpus, about 120 labeled queries
  • Recall@10: voyage-4-large 0.89, OpenAI 3-large 0.86, embed-v4.0 0.85
  • Cohere's clearest edge was the non-English subset

User Reports

  • An AWS Marketplace reviewer found Cohere beat Amazon Titan on a PDF chatbot
  • G2 lists no reviews for Cohere Embed itself

How It Compares

Feature Cohere Embed 5 Voyage AI OpenAI 3-large Jina AI
Max input 128K tokens 32K tokens 8,192 tokens 32K tokens (small models)
Dimensions 256 to 2,048 256 to 2,048 Up to 3,072 Matryoshka; 1,024 on omni-small
Images Yes Multimodal model No Yes (omni models)
Public price per 1M Not published $0.12 (4-large) $0.13 Token packages, not listed
Private deployment VPC, on-prem Via MongoDB No On-prem license via Elastic

Sources: Cohere models · Voyage docs · OpenAI pricing · Jina embeddings

Is this your tool? Add a badge