Rerank model comparison

Production stacks mix cross-encoders, listwise models, late-interaction, and instruction-following rerankers. The headline of 2026 is that size stopped predicting quality: a 0.6B listwise model (Jina v3.5) edges Qwen3-Reranker-4B on BEIR in Jina's own test, and in Qwen's own tests the 4B beats its 8B on English retrieval. Architecture, latency, languages and cost below — with honest footnotes wherever score protocols differ.

Rerank model comparison: architecture, type, best-for, languages, latency, pricing and BEIR score for each model
Model Architecture Type Best for Languages Typ. latency 50 docs Pricing BEIR NDCG@10 Demo
Qwen3-Reranker-4B Cross-encoder Apache 2.0Best Qwen3 size Best English score of the three; instruction-aware (GPU) 100+ langs · 32K ctx GPU required Free (self-host) MTEB-R 69.76* —
Qwen3-Reranker-8B Cross-encoder Apache 2.0 Largest Qwen3 — leads on Chinese, multilingual, long-document and code; 4B edges it on English 100+ langs · 32K ctx GPU / multi-GPU Free (self-host) MTEB-R 69.02* —
Qwen3-Reranker-0.6B Cross-encoder Apache 2.0 Lightest Qwen3; still needs GPU for comfort 100+ langs · 32K ctx GPU recommended Free (self-host) MTEB-R 65.80* —
bge-reranker-v2-m3 Cross-encoder Open weights Proven self-host default; CPU-friendly 100+ langs CPU-capable; GPU faster Free (self-host) 55.36* —
mxbai-rerank-large-v1 Cross-encoder Open weights
Browser ✓		 Legacy generation; xsmall still powers the browser demo English CPU-capable; GPU faster Free (self-host) 49.32*
mxbai-rerank-large-v2 Cross-encoder Open weightsMultilingual Strongest mxbai generation; multilingual incl. Chinese 100+ langs GPU recommended Free (self-host) 57.49* —
mxbai-rerank-base-v2 Cross-encoder Open weightsMultilingual Best size/quality balance in the v2 generation 100+ langs CPU-capable; GPU faster Free (self-host) 55.57* —
Jina Reranker v3.5 Listwise Hosted APIWeights: non-commercial Jina's 2026 flagship; drop-in for v3, 1.22–1.56× faster 100+ langs API or GPU self-host Free trial tokens, then token packages 63.20* —
Jina Reranker v3 Listwise Hosted APIWeights: non-commercial 0.6B listwise; 64 docs in one 131K ctx; superseded by v3.5 100+ langs API or GPU self-host Free trial tokens, then token packages 62.10* —
jina-reranker v1 tiny Cross-encoder Open weights
Browser ✓		 Browser / edge; powers our demo English ~229 ms per 10 passages (browser) Free (self-host) 48.54*
Cohere Rerank 4 Pro Cross-encoder Hosted API Mature multilingual API; 32k ctx 100+ langs API $0.0025/search† not published* —
Cohere Rerank 4 Fast Cross-encoder Hosted API Throughput / latency-tuned sibling 100+ langs API $0.002/search† not published* —
Voyage rerank-3 Cross-encoder Hosted API Voyage’s current flagship; 32k ctx Multilingual API $0.05/M tokens† (200M free) not published* —
Voyage rerank-3-lite Cross-encoder Hosted API Faster, cheaper current tier; 32k ctx Multilingual API $0.02/M tokens† (200M free) not published* —
Voyage rerank-2.5 Cross-encoder Hosted APIInstruction Older; the one documented for instructions; 32k ctx Multilingual API $0.05/M tokens† not published* —
Voyage rerank-2.5-lite Cross-encoder Hosted API Older cheaper tier; 32k ctx Multilingual API $0.02/M tokens† not published* —
Alibaba Cloud Model Studio (qwen3-rerank) Reranker API Hosted API Alibaba Cloud’s hosted Qwen rerankers: qwen3-rerank (500 documents, 4K tokens each), qwen3.7-text-rerank (30K tokens each, Beijing region only) and qwen3-vl-rerank for images and video 100+ langs · qwen3.7: 201 API ¥0.5/M tokens in China, $0.1/M internationally† not published* —
SiliconFlow (hosted open models) Reranker API Hosted API Hosted bge-reranker-v2-m3, bce-reranker-base_v1 and Qwen3-Reranker 0.6B / 4B / 8B behind one rerank endpoint Model-dependent API bge-reranker-v2-m3 free; Pro tier ¥0.07/M tokens† not published* —
Google Vertex AI ranking (semantic-ranker) Reranker API Hosted API Google Cloud’s ranking API; 1,024 tokens per record; semantic-ranker-default-004 is the default, the 005 models are in preview since Sep 2026 25 langs API $1.00 per 1,000 queries† not published* —
gte-reranker-modernbert-base Cross-encoder Open weightsCompact 149M, 8K context; also runs in transformers.js English Fast GPU / CPU-friendly Free (self-host) 56.19* —
gte-multilingual-reranker-base Cross-encoder Open weightsCompact 306M, 8K context; small multilingual default, served by Text Embeddings Inference 70+ langs Fast GPU / CPU-friendly Free (self-host) not published* —
Qwen3-VL-Reranker Cross-encoder Open weightsApache 2.0 Text, images, screenshots and video in one model; 2B / 8B, 32K context 30+ langs GPU Free (self-host) not published* —
zerank-2 Cross-encoder Open weightsApache 2.0 4B on Qwen3-4B, 32K context; also served by ZeroEntropy’s API. Relicensed to Apache 2.0 on Hugging Face in July 2026 — ZeroEntropy’s own docs still call it non-commercial English GPU Free (self-host) / API not published* —
llama-nemotron-rerank-1b-v2 Cross-encoder Open + NIM API 1.2B, 8K context; licensed for commercial use. NIM also serves a smaller llama-nemotron-rerank-500m-v2 Multilingual (26 langs evaluated) GPU / NIM NIM / self-host not published* —
ms-marco MiniLM-L6 (browser) Cross-encoder Open weights
Browser ✓		 Classic baseline; smallest of the three demo models (23M) English ~276 ms per 10 passages (browser) Free (self-host) 48.64*
LightOn-rerank Cross-encoderListwise Open weightsApache 2.0 Text and document-page images in one model; 0.8B / 2B / 4B, pointwise or listwise English · French GPU Free (self-host) 49.13* —
Contextual AI Rerank v2 Instruction Open weightsWeights: non-commercial Instruction-following; 1B / 2B / 6B 100+ langs · 32K ctx GPU Free (non-commercial self-host) not published* —
ColBERTv2 Late-interaction Open weights Token MaxSim; stage-1.5 not full CE English Fast rescore @ scale Free (self-host) n/a —

Last verified: October 2026 (added Qwen3-VL-Reranker, zerank-2 and gte-multilingual-reranker-base; added jina-reranker-v3.5; Jina rows now share one 2026 protocol; bge-reranker-v2-m3 corrected from an untraceable ~60.1 to BAAI's own 55.36*; Jina v1 tiny and ms-marco MiniLM-L6 now show the BEIR figures Jina published for them; in September, mxbai-rerank-large-v1 was corrected from an unsourced ~62.1) · Next review: Nov 2026. * No two vendors run BEIR the same way — BAAI average 15 datasets over bge-large-en-v1.5's top 100, Jina 13 datasets over jina-embeddings-v5-text-small's top 100, mixedbread their own set — so every BEIR figure here carries a * and its protocol on hover. The differences are not small: Jina's run puts bge-reranker-v2-m3 at 56.51* against BAAI's own 55.36*, and mxbai-rerank-large-v2 at 61.44* against mixedbread's 57.49*. Read the column as rough tiers, not decimals. Other * cells use MTEB-R or task-specific metrics. Where a vendor publishes no comparable number we say not published rather than guess. † Note the differing units: Cohere bills per search (one query + up to 100 docs, each counted once per 500 tokens), Voyage per token. Our cost calculator works out which is cheaper for your volume. Demo-capable models only: mxbai xsmall, Jina tiny, ms-marco MiniLM.

Sources: Qwen3-Reranker · jina-reranker-v3.5 · jina-reranker-v3 paper · jina-reranker-v1 · Voyage rerankers · mxbai-rerank · nemotron-rerank · MTEB · BEIR · BAAI evals · GTE ModernBERT · mGTE reranker · Qwen3-VL-Reranker · zerank-2 · Qwen guide · instruction · ColBERT

Tagged as a reranker, but not one for RAG: CLM-v0.1-8B from Contrastive-LM (September 2026) was the most-liked text-ranking model on Hugging Face by early October. It scores candidate actions and answers for agents — two small heads on a frozen Qwen3-8B, which you serve separately for its embeddings — and its headline results are agent benchmarks with fine-tuned heads, not retrieval. That is why it isn’t in this table.

Qwen3-Reranker

Apache 2.0Multilingual

0.6B / 4B / 8B, 32K context, Apache 2.0. Start at 4B — in Qwen's own tests it beats the 8B on English and on instruction-following, and trails it by at most 1.5 points elsewhere.

bge-reranker

Open weightsSelf-host

Open-weight rerankers from BAAI. Still the best CPU-friendly default (v2-m3). Pair with Qwen3 when you have GPU headroom.

Cohere Rerank 4

Hosted APIMultilingual

Pro and Fast variants, 32k context, 100+ languages. Billed per search: one query plus up to 100 documents, each counted once per 500 tokens.

Jina Reranker v3.5

Hosted APIListwiseWeights: non-commercial

0.6B listwise model at 63.20* BEIR in Jina's own test, just ahead of Qwen3-Reranker-4B's 62.28* there — though the 4B still leads on multilingual and legal/medical retrieval. Self-hosting the weights commercially needs a licence. v1-tiny (Apache 2.0) still powers our demo.

Voyage rerank-3

Hosted APIPer token

rerank-3 and -lite are Voyage’s current rerankers: 32k context, priced per token, with 200M free tokens each. They keep rerank-2.5’s natural-language instructions, with the biggest gains on long documents and code in Voyage’s own tests.

mxbai-rerank

Open weightsBrowser xsmall

Apache 2.0 rerankers from mixedbread-ai. v1 (DeBERTa-v3) still has the browser-sized xsmall powering our demo; v2 (Qwen2.5-based) adds 100+ languages.

Late-interaction (ColBERT)

ArchitectureEmerging

When ColBERTv2 beats bi-encoders and when you still need a cross-encoder — decision guide for 2026 stacks.

Instruction-following rerank

EmergingAPI

When task instructions change relevance — Contextual AI-style rerankers vs plain cross-encoders.

How to choose

Open · GPU

Qwen3-Reranker

Best multilingual open weights — start at 4B

  • 0.6B / 4B / 8B, all Apache 2.0, all 32K context
  • 4B is the sweet spot — the 8B is twice the size for at most 1.5 points, and only outside English
  • Vendor numbers are MTEB-R, so verify on your own labelled set
Open · CPU

bge-reranker

Best free self-host when GPU is limited

  • Zero per-call cost — runs on your own infra
  • Strong multilingual quality (v2-m3 covers 100+ languages)
  • Drop-in with sentence-transformers, LangChain, LlamaIndex
API

Cohere Rerank 4

Easiest hosted API; pick Pro or Fast

  • Official SDK for Python, Node, Java, Go — one-liner integration
  • 32k context and 100+ languages on both variants
  • Pro for precision, Fast for throughput — $0.0025 vs $0.002 a search
Open + API

Jina Reranker v3.5

Top-tier BEIR from a model that fits on one GPU

  • 63.20* BEIR nDCG@10 from 0.6B in Jina's own test — level with Qwen3-Reranker-4B, ~7× its size
  • Listwise: the query and all candidates share one 131K-token context
  • Weights are CC BY-NC 4.0: use the API, or license them for commercial self-hosting
API

Voyage rerank-3

Per-token pricing that tracks what you send

  • rerank-3 for accuracy, rerank-3-lite for latency — $0.05 and $0.02 per 1M tokens
  • 200M free tokens per account on each; 32k context
  • Need the 33% Batch API discount? Voyage document it for the older rerank-2.5 only
Open

mxbai-rerank

Apache 2.0 open weights; v2 adds 100+ languages

  • Permissive Apache 2.0 licence — use commercially without restrictions
  • v1 xsmall variant runs in the browser (powers our live demo)
  • v2 (base/large) is multilingual, including Chinese — v1 was English-only

Try a cross-encoder live

See how any of these models would reorder your retrieval results — demo runs in your browser.

Open the demo →