Curiosity · AI Model
Cohere Embed v3
Cohere Embed v3 is a production-focused retrieval embedding model — it ships with input-type prompts (search_document, search_query, classification, clustering) so you get specialised behaviour per task without training multiple encoders. Embed-v3 multilingual covers 100+ languages and is the default on Cohere's RAG platform.
Model specs
- Vendor
- Cohere
- Family
- Embed 3
- Released
- 2023-11
- Context window
- 512 tokens
- Modalities
- text
- Input price
- $0.1/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Input-type prompts boost retrieval quality on asymmetric query/doc corpora
- Strong multilingual coverage — 100+ languages in the multilingual variant
- Robust on long-tail and noisy enterprise text
- Works natively with Cohere Rerank 3 for a tight two-stage pipeline
Limitations
- 512-token input is short versus OpenAI 3-large (8191) — chunk more aggressively
- Closed weights — not self-hostable outside Cohere Private Deployments
- 1024-dim output is fixed (no Matryoshka truncation)
Use cases
- Enterprise RAG over policy and HR documents
- Multilingual semantic search across 100+ languages
- Document clustering and topic discovery
- Classification and intent routing
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| BEIR (avg NDCG@10) | ≈55 | 2024 |
| MIRACL (18 languages) | ≈57 | 2024 |
Frequently asked questions
What is Cohere Embed v3?
Cohere Embed v3 is a family of embedding models (English and multilingual) designed for retrieval-heavy enterprise workloads. It introduced input-type prompts — tags such as search_document and search_query that configure the embedding for a specific downstream task.
What does the input_type parameter do?
Cohere Embed v3 exposes an input_type parameter with values like search_document, search_query, classification, and clustering. Setting it correctly aligns the produced vector with its use, improving retrieval NDCG compared with a single embedding space.
How many languages does Embed v3 multilingual support?
The multilingual variant covers more than 100 languages, including all major European and Indian languages, with competitive MIRACL retrieval scores.
How much does Cohere Embed v3 cost?
As of April 2026, Cohere Embed v3 is priced at roughly USD 0.10 per million input tokens on the Cohere API, with Bedrock and Azure mirroring the Cohere rate.
Sources
- Cohere — Embed v3 announcement — accessed 2026-04-20
- Cohere — Embed docs — accessed 2026-04-20