Curiosity · AI Model
Cohere Rerank 3 (Multilingual)
Cohere Rerank 3 (Multilingual), released April 2024, is Cohere's cross-encoder reranking model. It takes a query and a batch of candidate documents and returns relevance scores that reorder the candidates before they are fed to an LLM. Unlike dense-vector embedders, a reranker reads both query and document jointly, producing sharper top-k ordering. v3 supports 100+ languages, handles long documents (up to 4k tokens), and is tuned for JSON, code, and tabular content as well as prose. Offered via the Cohere API and AWS Bedrock.
Model specs
- Vendor
- Cohere
- Family
- Rerank
- Released
- 2024-04
- Context window
- 4,096 tokens
- Modalities
- text
- Input price
- $2/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Cross-encoder joint scoring is more accurate than dense similarity
- Handles 100+ languages
- Tuned for non-prose content (code, tables, JSON)
- Available on Cohere API and AWS Bedrock
Limitations
- Closed model — API only
- Higher per-request cost and latency than dense retrieval alone
- Max 4k-token documents per candidate
- Needs to be paired with a first-stage retriever
Use cases
- RAG top-k reranking after a retriever returns candidates
- Multilingual search over 100+ languages
- Reranking over code, tables, and JSON
- Enterprise search quality uplift
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| BEIR average NDCG@10 (over BM25) | large uplift, SOTA commercial reranker at release | 2024-04 |
| Multilingual Mr. TyDi | top-tier across 11 languages | 2024-04 |
Frequently asked questions
What is Cohere Rerank 3 (Multilingual)?
A cross-encoder reranking model that rescores retrieval candidates against a query, supporting 100+ languages and non-prose content like code and tables.
Why rerank on top of a dense retriever?
Dense retrievers are fast but coarse. A cross-encoder reranker reads query and document together, producing much sharper top-k ordering at the cost of extra latency.
How do I access Rerank 3?
Via the Cohere API (`/rerank` endpoint), AWS Bedrock, or OCI Generative AI.
Sources
- Cohere — Introducing Rerank 3 — accessed 2026-04-20
- Cohere Rerank API docs — accessed 2026-04-20