Cohere Embed V4

coherecohere/embed-v4

Multilingual multimodal embedding model capable of transforming images, texts, and interleaved content into vector representations. State-of-the-art performance with byte/binary quantization and matryoshka embeddings for compression.

Official model card ↗Release announcement ↗Cohere Blog ↗Amazon Bedrock availability ↗

Cohere Embed v4 (embed-v4.0) is Cohere's first production-grade multimodal embedding model, released April 15, 2025. Through TheRouter, it converts text and images into vectors in a shared space; each input item produces a separate vector. Fused text-image inputs and PDF documents are not supported by this endpoint. The model supports Matryoshka embeddings in four dimensions (256, 512, 1024, 1536 default) and a 128 k token context window — 256× longer than its V3 predecessor.

Embed v4 is fine-tuned for enterprise domains (finance, healthcare, manufacturing) and achieves state-of-the-art results on text-to-text, text-to-image, and text-to-mixed-modality retrieval. It supports over 100 languages for text and is available on the Cohere Platform, AWS Bedrock (since Oct 2025), SageMaker, Azure AI Foundry, and Oracle OCI. TheRouter exposes it as cohere/embed-v4; image inputs use the image_url object extension shown below.

Best for

• Visual retrieval from document pages converted to supported image files; submit each page as a separate image input
• Cross-lingual semantic search across 100+ languages in a single index; ideal for global knowledge bases and customer-support archives
• Industry-specific retrieval (finance, healthcare, manufacturing) where domain-tuned embeddings outperform generic models on specialized terminology
• Cost-sensitive vector storage with Matryoshka dimensions — drop from 1536 → 256 to cut storage ~83% while retaining most retrieval quality

Reach for something else if

• Pure code or math retrieval — specialized code embedding models (e.g. Qwen text-embedding-v4 or OpenAI text-embedding-3-large) remain stronger on code semantics
• Real-time image-only search at scale — models trained exclusively on vision (CLIP, SigLIP) may offer higher throughput for image-only workloads
• On-premise or air-gapped deployments — Embed v4 is only available through managed cloud endpoints (Cohere, AWS, Azure, OCI)

Context Length

Max Output

Input Priceper 1M tokens

$0.1296/ 1M tokens

Image input Priceper 1M tokens

$0.470/ 1M tokens

Modalities

textimage→embedding

Capabilities

Vision

Pricing Breakdown

Type	Rate
Input	$0.1296 / 1M tokens
Image input	$0.470 / 1M tokens

Image input: TheRouter platform price $0.47 per million image tokens, based on Cohere's historical rate; no additional margin applies to this image rate.

Supported Parameters

input_typedimensions

Specifications

Release date	2025-04-15Cohere Blog ↗	verified
Output dimensions	256, 512, 1024, 1536 (default)docs.cohere.com ↗	verified
Context length	128,000 tokensdocs.cohere.com ↗	verified
Multimodal input	Text and image as separate input itemsdocs.cohere.com ↗	verified
Supported languages (text)	100+docs.cohere.com ↗	verified
Image input language	English onlydocs.oracle.com ↗	verified
Training cutoff	Not publicly disclosed	unknown
Matryoshka embeddings	Yes (4 dimensions)docs.cohere.com ↗	verified

Benchmarks

Benchmark	Distribution	Score	Source
MTEB (Massive Text Embedding Benchmark) Pure-text score; multimodal capability not captured by standard MTEB		65.2score	Ailog RAG benchmark summary ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{"model":"cohere/embed-v4","input":["A sample document"]}'

API guide

Embeddings request

Embed text or images. Images use image_url objects, not plain URL strings; each input item yields a separate vector. Image tokens cost $0.47 per million at the TheRouter platform rate.

cURL

curl https://api.therouter.ai/v1/embeddings \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cohere/embed-v4",
    "input": [{ "type": "image_url", "image_url": { "url": "https://httpbin.org/image/png" } }],
    "input_type": "search_document",
    "dimensions": 256
  }'

More from cohere

cohere/command-a

Not in the current catalog; the description below is historical.

Cohere's flagship text-generation model; pair with Embed v4 for retrieval-augmented generation pipelines.

Browse the current model catalog

cohere/command-r-plus

Not in the current catalog; the description below is historical.

Enterprise-grade reasoning model; complements Embed v4 for complex post-retrieval synthesis.

Browse the current model catalog

Similar models

Cross-provider sibling models

amazon/titan-embed-v2

AWS-native text embedding with configurable dimensions (256-1024); lower max dimension than Embed v4 and text-only.

openai/text-embedding-3-large

Not in the current catalog; the description below is historical.

OpenAI's strongest general-purpose embedding model; 3072 dimensions, text-only, strong multilingual coverage but no native image support.

Browse the current model catalog

qwen/text-embedding-v4

Alibaba's latest embedding model with strong code and math retrieval; text-only, 8K context.

Fact ledger — every claim on this page traces here

source	URL	retrieved
Release date	Cohere Blog ↗	2026-05-25	verified
Output dimensions	docs.cohere.com ↗	2026-05-25	verified
Context length	docs.cohere.com ↗	2026-05-25	verified
Multimodal input	docs.cohere.com ↗	2026-05-25	verified
Supported languages (text)	docs.cohere.com ↗	2026-05-25	verified
Image input language	docs.oracle.com ↗	2026-05-25	verified
Training cutoff	—	—	unknown
Matryoshka embeddings	docs.cohere.com ↗	2026-05-25	verified
MTEB (Massive Text Embedding Benchmark)	Ailog RAG benchmark summary ↗	2026-05-25	to verify