Best Rerankers for RAG in 2026
A reranker re-scores your first-pass retrieval results so the most relevant ones reach the LLM. We compared the leading 2026 rerankers — managed APIs and open-weight cross-encoders — on relevance lift, latency, license, and language and modality coverage.
Mixpeek runs reranking as a stage in a managed multimodal retriever — over video, images, audio, and documents in your own object storage, not just text. Add a rerank stage after dense or hybrid retrieval and search it through one API. Build starts at $25/mo for up to 1M vectors.
Rerank your own search results, freeQuick Answer
The best overall option in this category is Cohere Rerank 4, especially for teams that want the best managed reranking quality with zero infrastructure. The rankings below compare each tool by strengths, limitations, pricing, and fit for production use.
Cohere Rerank 4
Best for teams that want the best managed reranking quality with zero infrastructure.
jina-reranker-v3
Best for long-document rag that benefits from scoring candidates together, self-hosted or via api.
Qwen3-Reranker
Best for teams that want top open-weight reranking quality with full control and multilingual coverage.
Skip the comparison? Mixpeek runs rerankers for RAG on your own data: extraction, indexing, and search in one platform.
How We Evaluated
Evaluated by the Mixpeek engineering team, who build and operate multimodal retrieval infrastructure in production. Last tested July 2026; rankings re-checked when the market shifts, with pricing and claims verified against each vendor's public documentation.
Relevance Lift
How much the reranker improves ranking quality (nDCG@10, recall@k, MRR) over first-stage retrieval on real, diverse queries.
Latency at Depth
End-to-end reranking time at a realistic candidate depth (top-50 to top-100), since a reranker calls the model for every query-document pair.
License & Self-Hostability
Open weights vs API-only, license terms (Apache 2.0 vs commercial), and whether you can run it on your own GPU for data residency.
Language & Modality Coverage
Multilingual support, context length, and whether the model handles more than plain text (long documents, code, cross-lingual pairs).
Quick answer
The short version, before the detail:
- Cohere Rerank 4best for teams that want the best managed reranking quality with zero infrastructure— The most widely-adopted managed reranker, with broad language coverage and Pro/Fast tiers in one API
- jina-reranker-v3best for long-document rag that benefits from scoring candidates together, self-hosted or via api— Listwise reranking of up to 64 docs in a 131k-token window at only ~0.6B parameters
- Qwen3-Rerankerbest for teams that want top open-weight reranking quality with full control and multilingual coverage— Apache-2.0 open weights across sizes with 100+ languages and vision-language variants
- BGE Reranker v2-m3best for the default self-hosted reranker baseline for most rag pipelines— The most widely-deployed open-source cross-encoder — cheap, multilingual, and battle-tested
- Voyage rerank-2.5best for production rag that needs high reranking quality with low added latency— One of the best quality-per-millisecond managed rerankers, tuned to keep latency low
- mixedbread mxbai-rerank-v2best for latency-sensitive self-hosted reranking on a permissive license— A tiny, modern Apache-2.0 cross-encoder that reranks fast on modest hardware
- NVIDIA Llama NeMoRetriever Rerankbest for enterprises running rag on nvidia infrastructure at high throughput— GPU-optimized reranking as a NIM microservice inside the NeMo Retriever stack
- ZeroEntropy zerankbest for teams chasing the top of the relevance leaderboard willing to trial a newer vendor— Tops several independent 2026 head-to-head reranker leaderboards on relevance
Overview
Best Rerankers for RAG: comparison at a glance
| # | Tool | Best for | Pricing | Key differentiator | Main limit |
|---|---|---|---|---|---|
| 1 | Cohere Rerank 4 | Teams that want the best managed reranking quality with zero infrastructure | Managed API, priced per search (roughly $0.001 on v3.5 up to ~$0.0025 on Rerank 4 Pro); free trial tier | The most widely-adopted managed reranker, with broad language coverage and Pro/Fast tiers in one API | API-only — no self-hosting, your candidates leave your environment |
| 2 | jina-reranker-v3 | Long-document RAG that benefits from scoring candidates together, self-hosted or via API | Open weights on Hugging Face; Jina hosted API metered per token with a free starter allowance | Listwise reranking of up to 64 docs in a 131k-token window at only ~0.6B parameters | Listwise inference is heavier than a plain cross-encoder at large candidate depth |
| 3 | Qwen3-Reranker | Teams that want top open-weight reranking quality with full control and multilingual coverage | Free, open weights (Apache 2.0); you pay for your own GPU inference | Apache-2.0 open weights across sizes with 100+ languages and vision-language variants | Larger sizes need real GPU memory to serve at low latency |
| 4 | BGE Reranker v2-m3 | The default self-hosted reranker baseline for most RAG pipelines | Free, open weights (Apache 2.0); self-hosted GPU or CPU infrastructure costs only | The most widely-deployed open-source cross-encoder — cheap, multilingual, and battle-tested | Newer listwise and LLM rerankers can beat it on the hardest queries |
| 5 | Voyage rerank-2.5 | Production RAG that needs high reranking quality with low added latency | Managed API, usage-based per-token pricing on the Voyage pricing page; free tier to start | One of the best quality-per-millisecond managed rerankers, tuned to keep latency low | API-only; candidates leave your environment |
| 6 | mixedbread mxbai-rerank-v2 | Latency-sensitive self-hosted reranking on a permissive license | Free, open weights (Apache 2.0); mixedbread also offers a hosted API priced by token volume | A tiny, modern Apache-2.0 cross-encoder that reranks fast on modest hardware | Smaller than the largest rerankers, so tops out below them on the hardest queries |
| 7 | NVIDIA Llama NeMoRetriever Rerank | Enterprises running RAG on NVIDIA infrastructure at high throughput | Model weights on Hugging Face; served via NVIDIA NIM (NVIDIA AI Enterprise licensing for production) | GPU-optimized reranking as a NIM microservice inside the NeMo Retriever stack | Best value assumes NVIDIA GPUs and the NIM/NeMo stack |
| 8 | ZeroEntropy zerank | Teams chasing the top of the relevance leaderboard willing to trial a newer vendor | Hosted API (usage-based); a smaller open-weight model is available for self-hosting | Tops several independent 2026 head-to-head reranker leaderboards on relevance | Newer vendor with a shorter production track record |
| 9 | Mixpeek | Teams that need reranking inside a multimodal retrieval pipeline over their own object storage | Reranking is part of the managed retriever pipeline; usage-based, Build from $25/mo for up to 1M vectors | Reranking as a stage in an agent-native multimodal retriever, not a text-only reranker | A retrieval platform, not a drop-in reranker model you host yourself |
Cohere Rerank 4
The managed reranking standard: a single API call that re-scores your candidate documents with strong multilingual quality. Rerank 4 ships in Pro (accuracy) and Fast (latency) variants, and the older v3.5 remains a cheaper option.
The most widely-adopted managed reranker, with broad language coverage and Pro/Fast tiers in one API
Strengths
- +One API call, always current, nothing to host
- +Broad multilingual and long-document support
- +Pro/Fast variants trade accuracy against latency
- +Widely integrated across RAG frameworks (LangChain, LlamaIndex, Haystack)
Limitations
- -API-only — no self-hosting, your candidates leave your environment
- -Per-search cost adds up at high query volume
- -Closed weights; you cannot fine-tune on your own relevance data
Real-World Use Cases
- •Adding a reranking stage to a multilingual customer-support RAG system without provisioning any inference infrastructure
- •Improving precision on a legal-search product where the top-3 results must be exactly right
Choose This When
When you want proven reranking quality as a single API call, cover many languages, and prefer opex over running GPUs.
Skip This If
When you need to self-host for data residency, want to fine-tune on your own labels, or query volume makes per-search pricing expensive.
jina-reranker-v3
A compact (~0.6B) listwise reranker that reads up to 64 candidate documents together in a 131k-token context and scores them jointly, rather than one query-document pair at a time. Strong BEIR results for its size, available as open weights and a hosted API.
Listwise reranking of up to 64 docs in a 131k-token window at only ~0.6B parameters
Strengths
- +Listwise scoring across up to 64 documents at once
- +Very long 131k-token context for long documents
- +Small enough to self-host on a single modern GPU
- +Open weights plus a metered hosted API
Limitations
- -Listwise inference is heavier than a plain cross-encoder at large candidate depth
- -Newer model with a shorter production track record than Cohere or BGE
- -Best quality needs a GPU; CPU inference is slow
Real-World Use Cases
- •Reranking long research-paper chunks where cross-document comparison improves the final ordering
- •Self-hosting a capable reranker on a single GPU to keep document text inside your own environment
Choose This When
When your documents are long, you want listwise quality, and you value the option to self-host a small model or call an API.
Skip This If
When you need the absolute lowest latency at very large candidate depths, where a lighter pairwise cross-encoder is faster.
Qwen3-Reranker
Alibaba's open-weight reranker family (available in multiple sizes, e.g. 4B and 8B, plus vision-language reranker variants) under Apache 2.0. Supports 100+ languages and a 32k context, and tops several open reranker leaderboards.
Apache-2.0 open weights across sizes with 100+ languages and vision-language variants
Strengths
- +Apache 2.0 open weights, fully self-hostable and fine-tunable
- +100+ languages and a 32k context window
- +Multiple sizes to trade quality against latency and cost
- +Vision-language reranker variants for multimodal candidates
Limitations
- -Larger sizes need real GPU memory to serve at low latency
- -You own the serving, scaling, and updates
- -Smaller sizes trade some quality for speed
Real-World Use Cases
- •Self-hosting a multilingual reranker under a permissive license for a data-residency-sensitive deployment
- •Fine-tuning a reranker on in-domain relevance labels to beat a generic API on your own corpus
Choose This When
When you want the strongest open-weight quality, need to self-host or fine-tune, and cover many languages or multimodal candidates.
Skip This If
When you have no GPU capacity or want a fully managed, zero-ops reranker.
BGE Reranker v2-m3
The open-source cross-encoder that became the default self-hosted reranker: multilingual, Apache 2.0, and small enough to serve cheaply. A dependable baseline that most RAG stacks reach for first when self-hosting.
The most widely-deployed open-source cross-encoder — cheap, multilingual, and battle-tested
Strengths
- +Apache 2.0, fully self-hostable and free
- +Multilingual (m3) with solid quality for its size
- +Small and fast — cheap to serve at scale
- +Huge community adoption and framework support
Limitations
- -Newer listwise and LLM rerankers can beat it on the hardest queries
- -You own serving and scaling
- -Fixed model — quality gains require swapping models or fine-tuning
Real-World Use Cases
- •Adding a low-cost reranking stage to a self-hosted RAG system serving high query volume
- •Establishing a free open-source reranking baseline before deciding whether a paid API is worth it
Choose This When
When you want a proven, free, multilingual reranker you can self-host cheaply and swap in without vendor lock-in.
Skip This If
When you need the absolute top of the quality leaderboard or listwise scoring across many long documents.
Voyage rerank-2.5
Voyage AI's managed reranker, tuned for a strong quality-to-latency balance. A common pairing with Voyage embeddings, and frequently cited as one of the most production-balanced API rerankers in 2026.
One of the best quality-per-millisecond managed rerankers, tuned to keep latency low
Strengths
- +Strong relevance with notably low latency for an API reranker
- +Pairs cleanly with Voyage embeddings for an all-in-one retrieval stack
- +Managed — no infrastructure to run
- +Good long-context handling
Limitations
- -API-only; candidates leave your environment
- -Usage-based token pricing scales with volume
- -Closed weights; no fine-tuning on your data
Real-World Use Cases
- •Reranking in a latency-sensitive chat product where every 100ms of added retrieval time is visible to users
- •Standardizing on one vendor for both embeddings and reranking to simplify the retrieval stack
Choose This When
When latency matters as much as quality, and you want a managed reranker that pairs with a strong embedding model.
Skip This If
When you must self-host or want to avoid per-token API costs at very high volume.
mixedbread mxbai-rerank-v2
mixedbread's open-weight cross-encoder reranker family (base ~0.5B, large ~1.5B) built on the Qwen-2.5 architecture, under Apache 2.0. A modern, compact self-hosted option with a hosted API available.
A tiny, modern Apache-2.0 cross-encoder that reranks fast on modest hardware
Strengths
- +Apache 2.0 open weights in small, fast sizes
- +Modern Qwen-2.5-based architecture with strong quality for its size
- +Self-host for free or call the mixedbread API
- +Low latency — the base model is tiny
Limitations
- -Smaller than the largest rerankers, so tops out below them on the hardest queries
- -Younger project than BGE or Cohere
- -Self-hosting means you own serving and scaling
Real-World Use Cases
- •Running a reranker on constrained hardware where a 0.5B model keeps latency and cost minimal
- •Adding reranking to an edge or on-prem deployment that requires an open, permissive license
Choose This When
When you want a small, fast, permissively-licensed reranker to self-host, and can trade a little top-end quality for speed.
Skip This If
When you need the highest possible relevance on difficult, long-document queries.
NVIDIA Llama NeMoRetriever Rerank
NVIDIA's enterprise reranker, served as a NIM microservice and optimized for GPU-accelerated, high-throughput deployment. Part of the NeMo Retriever stack for enterprises standardizing RAG on NVIDIA infrastructure.
GPU-optimized reranking as a NIM microservice inside the NeMo Retriever stack
Strengths
- +Optimized for high-throughput GPU serving via NIM
- +Fits enterprises already standardized on NVIDIA AI Enterprise
- +Model weights available on Hugging Face
- +Strong multilingual and long-context support
Limitations
- -Best value assumes NVIDIA GPUs and the NIM/NeMo stack
- -Heavier operational setup than a single API call
- -Enterprise licensing for the managed NIM path
Real-World Use Cases
- •Standardizing an enterprise RAG platform's retrieval stack on NVIDIA NeMo Retriever end to end
- •Serving reranking at very high throughput on existing GPU clusters
Choose This When
When you already run NVIDIA GPUs and NIM/NeMo and want reranking optimized for that stack at scale.
Skip This If
When you want a simple hosted API or have no NVIDIA infrastructure commitment.
ZeroEntropy zerank
A newer reranker line that has topped several independent 2026 reranker leaderboards on head-to-head relevance. Offered as a hosted API, with a smaller open-weight variant available for self-hosting.
Tops several independent 2026 head-to-head reranker leaderboards on relevance
Strengths
- +Leads some independent 2026 relevance leaderboards
- +Hosted API plus a smaller open-weight option
- +Focused specifically on ranking quality
- +Simple to trial against an incumbent reranker
Limitations
- -Newer vendor with a shorter production track record
- -Top model is API-first; the open variant is smaller
- -Independent benchmarks vary by corpus — validate on your own data
Real-World Use Cases
- •Running a bake-off where the top-of-leaderboard reranker is one of the candidates on your own queries
- •Upgrading an existing reranker when a measurable relevance gain justifies switching vendors
Choose This When
When maximum measured relevance is the priority and you will benchmark candidates on your own corpus before committing.
Skip This If
When you need a long production track record or a fully open-weight top model.
Rather than a standalone reranker, Mixpeek runs reranking as a configurable stage inside a managed multimodal retriever. First-stage dense, sparse, or hybrid retrieval feeds a rerank stage, and the whole pipeline runs over your video, images, audio, and documents in your own object storage — not just text.
Reranking as a stage in an agent-native multimodal retriever, not a text-only reranker
Add a rerank stage after dense or hybrid retrieval in an MVS retriever, then query video, image, audio, and document results through one API.
Strengths
- +Reranking as one stage in a full multimodal retrieval pipeline
- +Works over video, images, audio, and documents, not text alone
- +Runs on your own object storage (S3/GCS/B2) with agent-native APIs
- +Combine first-stage hybrid retrieval and reranking without stitching services
Limitations
- -A retrieval platform, not a drop-in reranker model you host yourself
- -Reranker model choice is managed within the pipeline rather than fully arbitrary
- -Most valuable when you need multimodal retrieval, not text-only reranking
Real-World Use Cases
- •Reranking mixed video, document, and image results for an agent that searches an entire media library
- •Adding a rerank stage to a hybrid (BM25 + dense) retriever without running separate reranking infrastructure
Choose This When
When your retrieval spans modalities and you want first-stage search and reranking configured together over object storage.
Skip This If
When you only need to rerank text passages and already have a first-stage retriever — a standalone reranker model is simpler.
Which one should you choose?
- Choose Cohere Rerank 4 when you want proven reranking quality as a single API call, cover many languages, and prefer opex over running GPUs.
- Choose jina-reranker-v3 when your documents are long, you want listwise quality, and you value the option to self-host a small model or call an API.
- Choose Qwen3-Reranker when you want the strongest open-weight quality, need to self-host or fine-tune, and cover many languages or multimodal candidates.
- Choose BGE Reranker v2-m3 when you want a proven, free, multilingual reranker you can self-host cheaply and swap in without vendor lock-in.
- Choose Voyage rerank-2.5 when latency matters as much as quality, and you want a managed reranker that pairs with a strong embedding model.
- Choose mixedbread mxbai-rerank-v2 when you want a small, fast, permissively-licensed reranker to self-host, and can trade a little top-end quality for speed.
- Choose NVIDIA Llama NeMoRetriever Rerank when you already run NVIDIA GPUs and NIM/NeMo and want reranking optimized for that stack at scale.
- Choose ZeroEntropy zerank when maximum measured relevance is the priority and you will benchmark candidates on your own corpus before committing.
Put rerankers for RAG to work
Connect a bucket and Mixpeek runs the whole rerankers for RAG pipeline for you: extraction, indexing, and search over your own objects. No models to wire up, nothing to host.
Start with ManagedAlready have vectors?
Keep your embeddings on your own cloud and run dense, sparse, and BM25 search directly on object storage. From $25/mo.
Start with MVSFrequently Asked Questions
What is a reranker and why does it improve RAG?
A reranker is a second-stage model that re-scores the candidate documents returned by first-stage retrieval so the most relevant ones rise to the top before they reach the LLM. First-stage retrieval (dense vectors or BM25) is optimized for recall and returns many roughly-relevant candidates, but it compresses each document into a single vector computed without seeing the query. A reranker reads the query and each document together and assigns a precise relevance score, catching signals a single embedding misses. In practice this lifts answer quality because the LLM sees a cleaner, better-ordered context — often the single highest-ROI upgrade to a RAG pipeline after the embeddings themselves.
What is the difference between a cross-encoder and an LLM (listwise) reranker?
A cross-encoder scores one query-document pair at a time: it concatenates the query and a document, runs them through the model, and outputs a relevance score. This is accurate and reasonably fast, and models like BGE reranker v2-m3 and mxbai-rerank are cross-encoders. A listwise LLM reranker instead reads a batch of candidate documents together and orders them jointly, so it can compare documents against each other — jina-reranker-v3 and reasoning rerankers work this way. Listwise scoring often produces better ordering on hard queries because comparison is relative, but it costs more compute per query. A common pattern is a cheap cross-encoder for most traffic and a listwise reranker for the hardest queries.
Should I use an API reranker or self-host an open-weight model?
Use a managed API (Cohere, Voyage, Jina) when you want the best quality with zero infrastructure, cover many languages, and your query volume keeps per-search costs reasonable. Self-host an open-weight model (BGE reranker v2-m3, Qwen3-Reranker, mxbai-rerank-v2, all Apache 2.0) when you need data residency, want to fine-tune on your own relevance labels, or run enough volume that per-query API fees exceed the cost of a GPU. The break-even usually arrives faster than teams expect, because a small cross-encoder reranks cheaply on modest hardware. Whichever you pick, measure relevance lift on your own corpus rather than trusting a public benchmark.
How much latency does reranking add?
Reranking calls the model once for every candidate, so latency scales with candidate depth. Reranking the top 50-100 with a small cross-encoder typically adds tens to low-hundreds of milliseconds on a GPU; a large listwise model or a remote API can add more. The standard tactics to control it are keeping the first-stage candidate set tight (rerank the top 50, not the top 500), batching the pairs, serving the model on a GPU, and reserving the heaviest listwise rerankers for only the hardest queries. Because reranking usually improves precision enough to let you retrieve fewer candidates overall, the net latency cost is often smaller than it first appears.
Do rerankers work for multimodal (image, video, audio) retrieval?
Text rerankers only score text, so for multimodal retrieval you either rerank on a text projection of each result (transcripts, captions, OCR, extracted metadata) or use a vision-language reranker such as Qwen3's VL reranker variants that score image-text pairs directly. The cleaner approach is to make reranking a stage inside a multimodal retrieval pipeline: Mixpeek runs first-stage hybrid retrieval and a rerank stage over video, images, audio, and documents in your own object storage, so the same query is reranked consistently across modalities rather than bolting a text-only reranker onto multimodal results.
See how Mixpeek handles this
Purpose-built for rerankers for rag — not bolted on.
Talk to a Mixpeek engineer — free
30 minutes. Bring your use case and we'll tell you exactly what would work and what wouldn't.
Explore Other Curated Lists
Best Reverse Video Search Tools
Reverse video search finds where a clip appears, which videos are near-duplicates, and which library footage is visually similar to a query video. We compared the leading tools on match accuracy, clip and frame-level granularity, index scale, and how they handle re-encodes, crops, and edits.
Best Video Search Tools
We tested the leading video search and understanding platforms on real-world content libraries. This guide covers visual search, scene detection, transcript-based retrieval, and action recognition.
Best Visual Search APIs
A comparison of APIs that enable search-by-image functionality for ecommerce, stock photography, and visual asset management. We tested with real product catalogs and image libraries.