# Cohere Rerank 4 Fast

Cohere Rerank 4 Fast is the latency-optimized variant in the Cohere Rerank 4 family, released December 11, 2025. It reorders candidate documents by semantic relevance to a query across 100+ languages with a per-document context of 32K tokens.

- **Model ID:** `cohere/rerank-v4-fast`
- **Type:** reranking
- **Providers:** cohere
- **Context window:** 32,000
- **Maximum output tokens:** 32,000
- **Pricing:** See the provider details below
- **Canonical page:** https://vercel.com/ai-gateway/models/rerank-v4-fast

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { rerank } from 'ai';

const result = await rerank({
  model: 'cohere/rerank-v4-fast',
  query: 'What is the capital of France?',
  documents: [
    'Paris is the capital of France.',
    'Berlin is the capital of Germany.',
    'Madrid is the capital of Spain.',
  ],
})
```

## About

Cohere Rerank 4 Fast is the fast tier of Cohere's Rerank 4 generation, released December 11, 2025 alongside `rerank-v4-pro`. It shares the v4 family's multilingual coverage and JSON support but is tuned for lower per-query latency and higher throughput.

Like its siblings, Cohere Rerank 4 Fast is a cross-encoder. It reads the query and each candidate document together through attention, scoring relevance directly rather than relying on independent vector similarity. That structure picks up signal on complex, multi-part, or ambiguous queries that bi-encoder embeddings flatten.

Cohere Rerank 4 Fast covers more than 100 languages, including the same multilingual coverage as Cohere's embed-multilingual family. A French query can match Japanese documents and vice versa under one model. It handles long-form text, tables, code, and semi-structured JSON records with the same per-document context window of 32K tokens.

The fast variant fits at the front of high-traffic search systems and agentic retrieval steps where the reranker runs on every user turn. It pairs naturally with a multilingual embedding retriever for first-pass candidate selection and trades some quality versus `rerank-v4-pro` for response time under load.

See https://cohere.com/blog/rerank-4 for the request format, including `top_n`, `return_documents`, and per-document tokenization rules. Reranking is billed per search query rather than per token.

## What to consider

Rerankers run after an initial retrieval step. Pair Cohere Rerank 4 Fast with an embedding model, keyword search, or hybrid retriever so it has a candidate pool to score. The per-document context of 32K tokens covers query and document tokens together, which is wide enough to score long passages and chunked enterprise documents without truncation in most cases.

## When to use

### Best For

- Multilingual RAG pipelines that need fast reranking across 100+ languages with one model
- High-throughput search systems where reranker latency directly impacts user-facing response time
- Agentic workflows that run a retrieval-then-rerank step on every tool call or turn
- Mixed-format corpora combining long text, JSON records, tables, and code in a single index
- Cost-sensitive deployments where the per-query rerank price is the binding constraint

### Consider Alternatives When

- Maximum reranking quality matters more than latency, in which case `rerank-v4-pro` is the better fit
- Your corpus is English-only and `rerank-v3.5` already meets your accuracy bar at a lower price
- Your retrieval task only needs first-pass embedding similarity and reranking adds no measurable lift

## Best for

- **Multilingual reranking:** One model covers 100+ languages for cross-lingual retrieval
- **High-throughput search:** Lower latency keeps reranker overhead small under load
- **Agentic retrieval loops:** Tool calls that retrieve then rerank on every turn stay within tight time budgets
- **Mixed-format corpora:** Long text, JSON, tables, and code share one reranking stage
- **Cost-sensitive RAG:** A lower per-query price than the Pro variant at scale

## Consider alternatives

- **Maximum quality:** `rerank-v4-pro` targets state-of-the-art relevance on complex queries
- **English-only corpora:** `rerank-v3.5` covers English RAG at a lower price point
- **No second stage needed:** First-pass retrieval alone meets the accuracy bar
- **Image or multimodal retrieval:** Use a multimodal retrieval model instead

## Frequently asked questions

### How does Cohere Rerank 4 Fast differ from `rerank-v4-pro`?

Cohere Rerank 4 Fast is the latency-optimized variant in the Rerank 4 family, tuned for low-latency, high-throughput use cases. `rerank-v4-pro` targets the highest relevance quality on complex queries. Both share the same multilingual coverage and per-document context of 32K tokens.

### Which languages does Cohere Rerank 4 Fast support?

More than 100 languages, with the same multilingual coverage as Cohere's embed-multilingual family. Cross-lingual queries work in one call, so a query in one language can match documents in another.

### What document types can Cohere Rerank 4 Fast rerank?

Long-form text, semi-structured JSON, tables, code, and email-style records. The per-document context window is 32K tokens, shared between the query and the document.

### How is Cohere Rerank 4 Fast billed on AI Gateway?

Reranking is priced per search query rather than per token. See the pricing section on this page for the current per-query rate.

### Do I still need an embedding model with Cohere Rerank 4 Fast?

Yes, for the first-pass retrieval step. Cohere Rerank 4 Fast scores a candidate set against the query; it does not retrieve from the full corpus. A common pattern is to retrieve 50 to 200 candidates with an embedding model, then rerank them down to a top-k of 5 to 20.

### How does Cohere Rerank 4 Fast compare to `rerank-v3.5`?

`rerank-v3.5` targets English and matches `embed-multilingual-v3.0` coverage. Cohere Rerank 4 Fast is part of the Rerank 4 generation, explicitly multilingual across 100+ languages, and tuned for lower latency than `rerank-v4-pro`.

### Does Cohere Rerank 4 Fast support Zero Data Retention?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/rerank-v4-fast)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://cohere.com/blog/rerank-4)
- [Provider pricing](https://cohere.com/pricing)