jina-reranker-v3.5
by jinaai
A 0.6B multilingual listwise reranker that scores 63.20 nDCG@10 on BEIR, above a 4B Qwen reranker, under a non-commercial license
jinaai/jina-reranker-v3.5Deploy jina-reranker-v3.5
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
jina-reranker-v3.5 is Jina AI's multilingual document reranker: given a query and a list of candidates, it scores them all in one pass and returns them in order of relevance. It is a drop-in upgrade to jina-reranker-v3, built on Qwen3-0.6B.
On the card's top-100 protocol it scores 63.20 nDCG@10 on BEIR, ahead of v3 (62.10) and of Qwen3-Reranker-4B (62.28), with the largest gain on structured data (48.3 against 38.7). Qwen3-Reranker-4B still leads on MIRACL, RTEB and Struct-IR. The weights are CC-BY-NC-4.0, so commercial use needs Jina's API or a commercial agreement.
Architecture
Qwen3-0.6B with 28 layers in a hybrid attention pattern: sliding-window layers (window 1024) interleaved with global layers, ending on a global layer. The query and candidates go through together (listwise, "last but not late"), and a small MLP projector (1024 to 512 to 512) produces the per-document representation the scores come from.
Mixpeek SDK Integration
# A reranker scores text, so index documents (or the text attached to media)
# in a Mixpeek collection first; the reranker runs on what search returns.
import requests
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"blobs": [{"property": "body", "type": "text", "data": "Clause 14.2: either party may terminate on 30 days notice."}]},
)Capabilities
- Reranks a query against many candidate documents in one forward pass (listwise)
- Multilingual, including structured data such as tables and records
- Up to 131K tokens of query plus candidates
- GGUF and MLX builds for local use
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| BEIR | nDCG@10 | 63.20 | Model card: jinaai/jina-reranker-v3.5 (self-reported, top-100 from jina-embeddings-v5-text-small; v3 62.10, Qwen3-Reranker-4B 62.28) |
| MIRACL (multilingual) | nDCG@10 | 74.11 | Model card (self-reported; Qwen3-Reranker-4B 76.56) |
| RTEB | nDCG@10 | 70.95 | Model card (self-reported; Qwen3-Reranker-4B 77.68) |
| Struct-IR (structured data) | nDCG@10 | 48.3 | Model card (self-reported, controlled-pool protocol; v3 38.7) |
Performance
Non-commercial license: CC-BY-NC-4.0 rules out production use without a commercial agreement with Jina (its API, or the AWS, Azure and GCP listings the card links). The card reports 1.22x to 1.56x faster listwise inference than v3 on an A100. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
Can I use jina-reranker-v3.5 commercially?
Not the open weights. They are released under CC-BY-NC-4.0, which forbids commercial use. For production, use Jina's API or the cloud marketplace listings the card links, or choose a reranker with a permissive license.
How does jina-reranker-v3.5 compare with jina-reranker-v3?
It keeps v3's interface and size and improves on the card's benchmarks: BEIR 63.20 against 62.10, structured data 48.3 against 38.7, and better multilingual and legal retrieval. The card also reports 1.22x to 1.56x faster inference.
What is a listwise reranker?
A pointwise reranker scores each document against the query separately. A listwise reranker reads the query and all candidates together and ranks them in one pass, so it can compare candidates with each other and usually needs fewer model calls.
How do I use jina-reranker-v3.5 with Mixpeek?
Run the search with a Mixpeek retriever, then pass the returned documents' text to the reranker and keep its order. Check the license first: the open weights are non-commercial.
Specification
Research Paper
jina-reranker-v3.5: An Efficient Listwise Reranker with Hybrid Attention and Self-Distillation
arxiv.orgBuild a pipeline with jina-reranker-v3.5
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free