Which reranker should I use?

Tool ·

Four questions, and a starting point. The answers come from this site’s model reviews and guides; none of them replaces a test on your own queries.

Your situation

What are you ranking?

Where will it run?

Which languages?

Do you need to tell it what counts as relevant?

Start with

Every suggestion is listed below; with JavaScript on, the list narrows to the ones that match your answers.

A hosted API: Cohere Rerank 5 or Voyage rerank-3

Both bill per token, at the same list price. Cohere Rerank 5 returns calibrated scores, so one threshold works across queries. Voyage rerank-3 also takes instructions if you need them later. The cost calculator compares them with the other APIs at your volume.

An API that takes instructions: Voyage rerank-3 or Contextual AI

Voyage rerank-3 takes natural-language instructions; Cohere Rerank 5 does not. Contextual AI Rerank v2 is built around instructions and is sold as an API (its open weights are for non-commercial use). The instruction guide shows what to write.

In mainland China: Alibaba Cloud Model Studio or SiliconFlow

Alibaba Cloud Model Studio sells qwen3-rerank by the token in yuan, and SiliconFlow serves bge-reranker-v2-m3 and Qwen3-Reranker behind one endpoint. The mainland China guide compares them with hosting the weights yourself.

On your own GPU: Qwen3-Reranker-4B

Qwen3-Reranker covers 100+ languages and takes instructions. In Qwen’s own tests the 4B beats the 8B on English, so the largest model is not the automatic choice. The self-hosting guide has code that runs.

On CPU, multilingual: bge-reranker-v2-m3

bge-reranker-v2-m3 is the usual self-hosted default without a GPU. gte-multilingual-reranker-base is a smaller alternative that covers 70+ languages. The classic small models (ms-marco MiniLM, Jina v1, mxbai v1) were trained on English only.

On CPU, English: ms-marco MiniLM or jina-reranker-v1-tiny-en

A tiny cross-encoder is often enough when passages are short and latency matters; both run in the browser too, so you can try them in the demo first. For longer passages, look at the English GTE reranker.

Instructions on CPU are a stretch

The small CPU models above don’t take instructions. Qwen3-Reranker-0.6B does, but a GPU is recommended for it. If you can’t add one, an API that takes instructions is the simpler route.

Page images, self-hosted: Qwen3-VL-Reranker, LightOn-rerank or NVIDIA Nemotron

A text reranker only sees what OCR extracted; these score the page image itself, and all of them need a GPU. Qwen3-VL-Reranker and LightOn-rerank are Apache 2.0. NVIDIA’s Nemotron rerankers allow commercial use and can read a page’s image and its extracted text together.

Page images, hosted: Alibaba Cloud’s qwen3-vl-rerank

Alibaba Cloud Model Studio sells qwen3-vl-rerank as an API for images and video. The other visual rerankers on this site are open weights you host yourself, on a GPU: see the scenario guide.

For code, fix retrieval first

Vector search alone misses exact function names, and BM25 alone misses paraphrased questions. Use hybrid retrieval, then rerank. Qwen and Voyage both publish code-retrieval results for their rerankers.

Then check it

A suggestion is where to start, not the answer. Line the candidates up in the comparison guide or the models table, then evaluate two of them on fifty of your own queries. The scenario guide explains the reasoning behind each pick.