# Voyage Rerank 2.5 Lite

Voyage Rerank 2.5 Lite is Voyage AI by MongoDB's latency-and-quality balanced reranker. It supports instruction-following over natural language steering, multilingual retrieval, and a context window of 32K tokens, while remaining significantly smaller than the top open-source rerankers.

- **Model ID:** `voyage/rerank-2.5-lite`
- **Type:** reranking
- **Providers:** voyage
- **Context window:** 32,000
- **Maximum output tokens:** 32,000
- **Pricing:** $0.02/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/rerank-2.5-lite

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

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

const result = await rerank({
  model: 'voyage/rerank-2.5-lite',
  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

Voyage Rerank 2.5 Lite is Voyage AI by MongoDB's generalist reranker released August 11, 2025, optimized for both latency and quality. It reorders candidate documents returned by a first-stage retriever, with a context window of 32K tokens, multilingual support, and the same instruction-following capability as `rerank-2.5`.

Across 93 retrieval datasets, Voyage Rerank 2.5 Lite improves accuracy by 7.16% over Cohere Rerank v3.5 when paired with four first-stage retrieval methods: BM25 lexical search, OpenAI `text-embedding-3-large`, `voyage-3-large`, and `voyage-3.5`. Averaged across those first-stage methods, Voyage Rerank 2.5 Lite outperforms Cohere Rerank v3.5, Qwen3-Reranker-8B, and `rerank-2-lite` by 1.93%, 1.01%, and 2.70% respectively on NDCG@10. It performs better than Qwen3-Reranker-8B, the strongest open-source reranker in the comparison, despite being over an order of magnitude smaller.

On the Massive Instructed Retrieval (MAIR) benchmark, Voyage Rerank 2.5 Lite outperforms Cohere Rerank v3.5 by 10.36%. Instruction-following lets you steer relevance scores using natural language without changing your retrieval index. The context window of 32K tokens matches `rerank-2.5`, so long query-document pairs fit a single rerank call.

## What to consider

Voyage Rerank 2.5 Lite sits in the latency-and-quality balanced tier of the Voyage AI by MongoDB 2.5 reranker series. Pick it when per-query cost and response time matter and you can accept a small accuracy tradeoff versus `rerank-2.5`. Most production RAG pipelines fall in this category.

If retrieval quality is the dominant constraint, `rerank-2.5` is the quality-optimized tier of the same series. The two models share the same context window, instruction-following capability, and multilingual coverage.

Voyage Rerank 2.5 Lite pairs with any first-stage retriever. Keep your existing embedding stack and add Voyage Rerank 2.5 Lite as a second-stage reranker to lift top-k quality without changing how documents are indexed.

## When to use

### Best For

- **High-volume RAG pipelines** where per-query reranking cost scales with traffic
- **Latency-sensitive search** where customer-facing response time matters
- **Instruction-driven relevance** encoded in natural language without retraining
- **Multilingual reranking** across many languages in one model
- **Cost-sensitive production** that wants strong reranking without the top price tier

### Consider Alternatives When

- **Top-tier accuracy is the priority**; `rerank-2.5` is the quality-optimized tier
- **You need an open-source reranker** for self-hosting; Qwen3-Reranker-8B is an option, though Voyage Rerank 2.5 Lite outperforms it on Voyage's benchmarks
- **Your retrieval is single-language English-only** and a monolingual reranker covers the workload
- **A reranker stage may not be needed**; strong first-stage retrievers like `voyage-3.5` may meet your targets without a second pass

## Best for

- **High-volume RAG pipelines:** Per-query reranking cost scales with traffic, and Voyage Rerank 2.5 Lite balances accuracy with throughput
- **Latency-sensitive search:** Customer-facing search where end-to-end response time matters and a reranker stage must stay fast
- **Instruction-driven relevance:** Encode preferences such as recency or source authority in natural language without retraining
- **Multilingual reranking:** One reranker covers retrieval across many languages
- **Cost-sensitive production:** Strong reranking quality without the full price tier of `rerank-2.5`

## Consider alternatives

- **Top-tier accuracy is the priority:** `rerank-2.5` is the quality-optimized tier and improves on Cohere Rerank v3.5 by 7.94% versus 7.16% for Voyage Rerank 2.5 Lite
- **Open-source-only constraint:** Qwen3-Reranker-8B is available for self-hosting, though Voyage Rerank 2.5 Lite outperforms it on Voyage's published benchmarks
- **Single-language English-only retrieval:** A monolingual reranker may suffice when other languages are out of scope
- **No reranker stage needed:** Strong first-stage retrievers like `voyage-3.5` may meet your accuracy targets without a second pass

## Frequently asked questions

### What is the difference between Voyage Rerank 2.5 Lite and rerank-2.5?

`rerank-2.5` is the quality-optimized tier of the series and improves on Cohere Rerank v3.5 by 7.94%. Voyage Rerank 2.5 Lite is the latency-and-quality balanced tier and improves on Cohere Rerank v3.5 by 7.16%. Both support instruction-following, the same context window of 32K tokens, and multilingual retrieval.

### How does Voyage Rerank 2.5 Lite compare to Qwen3-Reranker-8B?

Voyage AI by MongoDB reports that Voyage Rerank 2.5 Lite performs better than Qwen3-Reranker-8B on the published benchmark suite despite being over an order of magnitude smaller. Averaged across four first-stage retrieval methods, Voyage Rerank 2.5 Lite outperforms Qwen3-Reranker-8B by 1.01% on NDCG@10.

### What is instruction-following in Voyage Rerank 2.5 Lite?

Instruction-following lets you steer relevance scores using natural language. You pass an instruction with the query and candidates, and Voyage Rerank 2.5 Lite adjusts scoring to reflect it. On the MAIR benchmark, this lifts Voyage Rerank 2.5 Lite 10.36% above Cohere Rerank v3.5.

### What is the context window for Voyage Rerank 2.5 Lite?

32K tokens. That is double the window of `rerank-2-lite` and eight times that of Cohere Rerank v3.5, so longer query-document pairs fit a single rerank call without truncation.

### Which first-stage retrievers does Voyage Rerank 2.5 Lite work with?

Voyage Rerank 2.5 Lite reorders candidates from any first-stage method. Voyage AI by MongoDB benchmarks it on BM25 lexical search, OpenAI `text-embedding-3-large`, `voyage-3-large`, and `voyage-3.5`. You can keep your existing embedding model and add Voyage Rerank 2.5 Lite as a second-stage reranker.

### Does Voyage Rerank 2.5 Lite support multilingual retrieval?

Yes. Voyage Rerank 2.5 Lite reranks across many languages without separate per-language models. Voyage AI by MongoDB reports consistent improvement across all evaluated languages and first-stage retrieval methods.

### How do I access Voyage Rerank 2.5 Lite through Vercel AI Gateway?

Add your Voyage AI by MongoDB API key in AI Gateway settings, then send rerank requests through AI Gateway. AI Gateway authenticates requests and records usage. You can call Voyage Rerank 2.5 Lite through the AI SDK alongside Chat Completions, Responses, and Messages API formats.

### Is Zero Data Retention available for Voyage Rerank 2.5 Lite?

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-2.5-lite)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.voyageai.com/docs/reranker)
- [Provider pricing](https://docs.voyageai.com/docs/pricing)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.