Best reranker in 2026: Qwen3, bge, mxbai, Jina, GTE and the hosted APIs compared

Comparison · ~7 min read ·

Every reranker vendor publishes a table in which its own model wins, and the tables use different datasets, so the numbers cannot be lined up across vendors. This page only sets scores side by side when they come from one run, says whose run it was, and says plainly which models have no number that can be compared.

The short answer

Which numbers can be compared

A score only means something next to scores from the same run: the same datasets, the same candidate lists from the same first-stage retriever, the same cut-off. Two numbers from two vendors’ pages are not on one scale, even when both say “nDCG@10” or “BEIR”. This page uses two runs that cover several models at once, and both were run by a vendor with a model of its own in the table, so read them as strong hints and confirm on your own labelled queries (how).

Three families have no run to put here. GTE: Alibaba publishes its reranking results as a chart, so there is no table to line up. Cohere: no benchmark table has been published for Rerank 5. hs-rerank-1: Hindsight’s model card shows its results as an image only. They are compared below on size, licence, price and fit instead.

Qwen3 against bge-reranker-v2-m3

This is Qwen’s own run, from the Qwen3-Reranker model cards: the retrieval subsets of MTEB (English, Chinese, multilingual and code) and MLDR for long documents, each reranking the same top 100 candidates.

ModelMTEB-RCMTEB-RMMTEB-RMLDRMTEB-Code
Qwen3-Reranker-8B69.02*77.4572.9470.1981.22
Qwen3-Reranker-4B69.76*75.9472.7469.9781.20
Qwen3-Reranker-0.6B65.80*71.3166.3667.2873.42
bge-reranker-v2-m357.0372.1658.3659.5141.38

The 4B and 8B lead bge-reranker-v2-m3 on every column, and the gap is widest on code. The 0.6B is the exception that matters if you work in Chinese: on CMTEB-R it scores 71.31 against bge’s 72.16, so shrinking to 0.6B costs the lead there. The 8B is not clearly better than the 4B: it is ahead only on Chinese, level on multilingual, long documents and code, and behind on English.

Qwen3, mxbai and Jina on one BEIR run

This is Jina’s 2026 evaluation, from the jina-reranker-v3.5 model card. Jina re-ran six models under one protocol (nDCG@10): 13 BEIR datasets, MIRACL for multilingual retrieval, RTEB for professional domains (legal, finance, code, medical) and Struct-IR for structured data, each reranking the top 100 candidates from jina-embeddings-v5-text-small. The table shows four of them; bge-reranker-v2-m3 was not in the run.

ModelBEIRMIRACLRTEBStruct-IR
jina-reranker-v3.5 (0.6B)63.2074.1170.9548.3
mxbai-rerank-large-v2 (1.5B)62.4569.6570.8143.0
Qwen3-Reranker-4B62.2876.5677.6855.6
Qwen3-Reranker-0.6B56.9467.1268.4141.9

On plain English BEIR the top three are within a point of each other, so that benchmark cannot pick between them. The differences show up elsewhere: Qwen3-Reranker-4B leads on multilingual, professional-domain and structured retrieval by clear margins. Jina’s own model tops BEIR in Jina’s own run; its weights are non-commercial, so for a product the practical comparison is Qwen3 against mxbai.

Size, licence and where each runs

ModelSizeLicenceLanguagesRun it on
Qwen3-Reranker0.6B / 4B / 8BApache 2.0100+GPU recommended
bge-reranker-v2-m3568MOpen weights100+CPU or GPU
mxbai-rerank-large-v21.5BApache 2.0100+GPU recommended
jina-reranker-v3.50.6BNon-commercial weightsMultilingualHosted API, or weights for research
hs-rerank-1114MMIT8, incl. Chinese, Japanese and KoreanCPU (int8 ONNX) or GPU
GTE rerankers149M / 306MApache 2.0English / 70+CPU, small GPU or browser (English)

Sizes and licences are from each model’s own page here, which cite the vendors’ model cards; hs-rerank-1 has no page of its own, so its row comes straight from its model card. “Open weights” means the licence is not restated on this page; check the card before shipping.

Hosted APIs

You pay per token instead of running a GPU. None of these vendors publishes a comparable benchmark, so the table gives what can be checked: the price and the context length.

APIPrice per million tokensContextBenchmark
Cohere Rerank 5 Pro / Fast$0.05 / $0.0232,768 tokensNone published
Voyage rerank-3 / lite$0.05 / $0.0232,000 tokensVoyage’s own evaluations only
zerank-2$0.02532,768 tokensZeroEntropy’s own run only

Per-token and per-search prices are billed differently, so the cheapest sticker price is not always the cheapest bill. The cost calculator works it out for your documents per query and queries per month.

How to decide

  1. Pick two candidates from this page: one open model you could host and one hosted API.
  2. Label 30–50 real queries from your own users with the passages that answer them.
  3. Run both and compare nDCG@10 on your labels, with the same candidate list. This is the only comparison that predicts your result; the evaluation guide walks through it.
  4. Check that reranking is the bottleneck. If the right passage is never in the candidates, no reranker helps; see why reranking didn’t help.

See the whole market in one table

Every model and hosted API, with the protocol behind each score, and a one-click jump to the browser demo.

Compare all models → Choose by scenario →

Keep reading