Rerank model comparison

Production stacks mix cross-encoders, listwise models, late-interaction, and instruction APIs. The headline of 2026 is that size stopped predicting quality: a 0.6B listwise model (Jina v3) outscores Qwen3-Reranker-4B on BEIR, a 149M cross-encoder ties a 1.2B one on Hit@1, and Qwen3's own 4B edges its 8B. 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 Strongest multilingual open self-host (GPU) 100+ langs · 32K ctx GPU required Free (self-host) MTEB-R* —
Qwen3-Reranker-8B Cross-encoder Apache 2.0 Largest Qwen3 — but 4B edges it on BEIR 100+ langs · 32K ctx GPU / multi-GPU Free (self-host) MTEB-R* —
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* —
bge-reranker-v2-m3 Cross-encoder Open weights Proven self-host default; CPU-friendly 100+ langs 200–500 ms CPU
20–50 ms GPU		 Free (self-host) 55.36* —
mxbai-rerank-large-v1 Cross-encoder Open weights
Browser ✓		 Legacy generation; xsmall still powers the browser demo English 150–400 ms CPU
~30 ms GPU		 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 Listwise Open + HostedBeats Qwen3-4B 0.6B listwise; 64 docs in one 131K ctx 100+ langs API or GPU self-host Free tier + pay-as-you-go 61.94* —
jina-reranker v1 tiny Cross-encoder Open weights
Browser ✓		 Browser / edge; powers our demo English 30–80 ms CPU (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-2.5 Cross-encoder Hosted APIInstruction Instruction-following; 32k ctx Multilingual API $0.05/M tokens† not published* —
Voyage rerank-2.5-lite Cross-encoder Hosted API Cheaper tier; same 32k ctx Multilingual API $0.02/M tokens† not published* —
Voyage rerank-3 Cross-encoder Hosted APIPreview Preview successor to 2.5; carries the free-token grant Multilingual API $0.05/M tokens† (200M free) not published* —
Voyage rerank-3-lite Cross-encoder Hosted APIPreview Cheaper preview tier; same 32k ctx Multilingual API $0.02/M tokens† (200M free) not published* —
gte-reranker-modernbert-base Cross-encoder Open weightsCompact ~149M, yet ties nemotron-1b on Hit@1 English Fast GPU / CPU-friendly Free (self-host) Hit@1 ≈ nemotron* —
llama-nemotron-rerank-1b-v2 Cross-encoder Open + NIM API 1.2B; top accuracy when latency is free Multilingual (MIRACL / MLQA evals) GPU / NIM NIM / self-host Hit@1 83.0* —
ms-marco MiniLM-L6 (browser) Cross-encoder Open weights
Browser ✓		 Classic baseline; largest of the three demo models English 50–150 ms browser WASM Free (self-host) 48.64*
Contextual AI Rerank v2 Instruction Hosted API Instruction-following / policy-shaped relevance English API Contact vendor† n/a —
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 (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-v3'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), 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 paper · jina-reranker-v1 · Voyage rerankers · mxbai-rerank · nemotron-rerank · MTEB · BEIR · BAAI evals · GTE ModernBERT · Qwen guide · instruction · ColBERT

Qwen3-Reranker

Apache 2.0Multilingual

0.6B / 4B / 8B, 32K context. Start at 4B — it reportedly edges the 8B on BEIR, so the largest size is not the automatic answer.

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 — not per document.

Jina Reranker v3

Open + HostedListwise

0.6B listwise model scoring 61.94 on BEIR — ahead of Qwen3-Reranker-4B at a sixth of the size. v1-tiny still powers our demo.

Voyage rerank-2.5

Hosted APIInstruction

rerank-2.5 and -lite, both 32k context with instruction following, so you can steer relevance in natural language. Priced per token.

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 costs more and scores no better
  • 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

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

  • 61.94 BEIR nDCG@10 from 0.6B — beats Qwen3-Reranker-4B, 6× its size
  • Listwise: 64 documents share one 131K-token context
  • v1-tiny still runs in the browser (our demo)
API

Voyage rerank-2.5

Instruction-following relevance, priced per token

  • Steer scoring with a natural-language instruction, no fine-tune
  • 32k context on both rerank-2.5 and the cheaper -lite tier
  • The preview rerank-3 tier now carries the 200M free-token grant that 2.5 launched with; 33% off via the Batch API
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 →