Back to Gap Map Model components / Embeddings & retrieval models

Cohere Embed

closed / Overall score: 3.7

Cohere’s embedding model line, currently embed-v4.0: a single multimodal model that embeds text, images and mixed text/image payloads such as PDFs into one space. It carries a 128k context, Matryoshka output dimensions of 256, 512, 1024 or 1536, and is sold through Cohere’s Embed endpoint, AWS SageMaker and Bedrock, and Azure AI Foundry.

embed-v4.0 is still the head of the line as of 2026-09-11 — Cohere’s own model table lists it as the latest, with the v3.0 English/multilingual pair below it and nothing newer. No weights are published for any SKU; the private-deployment option is a priced dedicated instance, not a download. Verified 2026-09-11 via the Cohere Embed model table, the Embed 4 release note and the Cohere pricing page.

Openness

1 high confidence

1.0

weights: closed(embed-v4.0 is served through the Embed endpoint and cloud marketplaces
data: closed(training corpus not released or described)
code: closed(no training or inference implementation
license: proprietary(hosted API

API-only, so the ladder's first rule fires. The dedicated-instance tier on the pricing page is worth naming because it reads like self-hosting and is not: Cohere rents an instance running the model, and nothing is downloadable.

https://docs.cohere.com/docs/cohere-embed recorded 2026-09-11
The Embed model table lists embed-v4.0 as the latest model with a 128k context and Matryoshka dimensions '[256, 512, 1024, 1536 (default)]', reached through the Embed endpoint. No weight download, self-hosted build is offered for it or for any v3.0 SKU. Two Hub repositories do exist under CohereLabs for the v3.0 embed models, but both ship only a config and a tokenizer - no weight file - so neither is a distribution of the model.
https://cohere.com/pricing recorded 2026-09-11
Prices Embed 4 as hosted usage and as dedicated instances — 'Embed 4 Small $4.00 hourly / $2,500 monthly per instance', 'Embed 4 Medium $5.00 / $3,250' — which is rented capacity, not a weights license.

Adoption

3 medium confidence

3.0

No artifact to count and no call volume published. The level rests on the same breadth of multi-cloud distribution that places cohere-rerank-api: Cohere's own platform, AWS SageMaker and Bedrock, and Azure AI Foundry, where embed-v-4-0 is sold directly by Microsoft.

https://docs.cohere.com/changelog/embed-multimodal-v4 recorded 2026-09-11
'Embed v4 is available today on the Cohere Platform, AWS Sagemaker, and Azure AI Foundry.' Carries no usage figure.
https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/models recorded 2026-09-11
Azure AI Foundry's models-sold-directly-by-Azure list carries embed-v-4-0, confirming Microsoft resells it.

Capability

4 high confidence

4.0

Rung 4, competitive frontier: current generation, multilingual, genuinely long-context at 128k, and multimodal in one model. Short of 5 because the only public head-to-head number is a rival's, and it puts Embed v4 8.20% behind the RTEB leader. Cohere's own SOTA claim is unquantified on the release note.

https://docs.cohere.com/changelog/embed-multimodal-v4 recorded 2026-09-11
'Embed v4 achieves state of the art in the following areas: Text-to-text retrieval, Text-to-image retrieval, Text-to-mixed modality retrieval (from e.g. PDFs)'; 'Context length of 128k'. No numeric result is given.
https://blog.voyageai.com/2026/01/15/voyage-4/ recorded 2026-09-11
Over all 29 RTEB datasets, voyage-4-large surpasses Cohere Embed v4 by an average of 8.20%.

Verified 2026-09-11