Skip to content

Jina Reranker v3.5

Small 0.6B multilingual reranker from Jina AI (now Elastic) that scores a whole candidate list in one pass, with 131K-token context.

Reviewed by Andrew Miller · Updated

Our rating

4.3/5

A 0.6B reranker that beats a 4B one on BEIR, held back by a licence that turns self-hosting into a sales call.

What it is

jina-reranker-v3.5 is a reranker from Jina AI, which Elastic acquired in October 2025; it was released on July 27, 2026. A reranker rescores a search system's shortlist by reading the query and documents together. Unlike most rerankers, which score one document at a time, this one is listwise: it reads the query and many candidates in a single pass, which lets documents be judged against each other.

It has 597 million parameters, is built on Qwen3-0.6B, was trained in 52 languages and accepts up to 131,072 tokens, which must hold the query and all candidates. It is a drop-in replacement for jina-reranker-v3: only the model name changes.

The weights are on Hugging Face under a non-commercial CC BY-NC 4.0 licence, with GGUF and MLX builds. Commercial use goes through the Jina API at $0.05 per million tokens, the Elastic Inference Service, AWS or Azure marketplace billing, or a paid Jina On-Prem licence from Elastic.

Our review

Jina Reranker v3.5 gets near-large-model ranking quality from a model small enough to run cheaply. On Jina's model card it scores 63.20 nDCG@10 on BEIR, ahead of Qwen3-Reranker-4B at 62.28 despite being about seven times smaller. Qwen's 4B model still wins on multilingual MIRACL (76.56 against 74.11), on the RTEB enterprise suite (77.68 against 70.95) and on structured data (55.6 against 48.3). These are Jina's own measurements, run with jina-embeddings-v5-text-small as the first stage.

The gains over v3 land where business search hurts. Elastic's launch post reports AILA case-law retrieval up from 20.82 to 32.55 and statute retrieval from 32.15 to 46.16, while LegalQuAD and LegalSum barely move. Medical and finance results appear only as a chart. Elastic is candid that a reranker cannot rescue bad first-stage retrieval.

Latency depends heavily on document length. Elastic measured 305 ms per request on BEIR Natural Questions, whose documents average 146 tokens, but 10.3 seconds on legal case files averaging 1,904 tokens, even after a 56% speed-up over v3. Long-document reranking belongs in offline or asynchronous paths unless candidates are trimmed first.

The licence decides most buying decisions. CC BY-NC 4.0 limits the open weights to testing and research, so production self-hosting requires Elastic's Jina On-Prem licence at a quoted annual price. The hosted API is cheap at $0.05 per million tokens, but there it meets Voyage's rerank-3 and Cohere's Rerank 4 head on, and Jina's case rests on its own benchmark table.

Where it earns its place

  • An Elasticsearch team running hybrid search over a help centre

    Lift answer quality without adding a second vendor

    Call the reranker through the Elastic Inference Service after BM25 and vector retrieval. If each request carries the top 50 chunks of about 300 tokens, that is roughly 15,000 tokens, or $0.00075 a query at $0.05 per million on the Jina API.

  • A legal-tech startup ranking case law for lawyers

    Put the right precedent in the top five results

    v3.5's biggest published gains are on AILA case and statute retrieval. Long judgements make requests slow, about 10 seconds at 1,904-token documents in Elastic's test, so rerank passages rather than full files.

  • A RAG platform engineer serving 1M queries a month

    Add reranking at a predictable cost and throughput

    At about 15,000 tokens a request, 1M queries cost around $750 a month. The paid tier's 2M tokens a minute caps that at roughly 130 requests a minute, so peak traffic needs a Premium key at 50M.

Before you commit

Setup
Swap the model name to jina-reranker-v3.5 in an existing v3 call, or load the weights in Transformers with trust_remote_code and call model.rerank.
Integrations
Jina API, Elastic Inference API and Service, AWS SageMaker and Azure billing, plus a Jina On-Prem container; GGUF and MLX builds for local evaluation. Text only; Jina points to jina-reranker-m0 for images.
Data and privacy
Jina says it does not train on customer inputs and stores them only as needed to serve requests, under Elastic's DPA; no retention period or certification is named. The weights' CC BY-NC 4.0 licence bars commercial self-hosting without Jina On-Prem.
Cost at scale
$0.05 per million tokens, so 1M queries of about 15,000 tokens each cost around $750 a month. Jina's page does not say whether the query is counted once or per document.
Limits
131,072 tokens per request across query and candidates, auto-truncated on the API. Rate limits are 100 requests and 100K tokens a minute free, 500 and 2M paid, 5,000 and 50M Premium.

What's good

  • Beats Qwen3-Reranker-4B on BEIR at 0.6B
  • 131K-token listwise context with no fixed document count
  • AILA case-law retrieval up from 20.82 to 32.55 over v3
  • Drop-in for v3; GGUF and MLX builds for local tests

What's not

  • Non-commercial licence on the weights
  • Trails Qwen3-Reranker-4B on MIRACL, RTEB and structured data
  • About 10 seconds per request on 1,900-token legal documents

Best for

Elastic users and teams that want a small, long-context reranker on their own hardware.

Alternatives

Jina Reranker v3.5: common questions

Is Jina Reranker v3.5 worth it?
A 0.6B reranker that beats a 4B one on BEIR, held back by a licence that turns self-hosting into a sales call. We rate it 4.3 out of 5.
How much does Jina Reranker v3.5 cost?
Jina API $0.05 per 1M tokens. Weights free for non-commercial use; commercial self-hosting needs a paid Elastic licence.
What are the downsides of Jina Reranker v3.5?
Non-commercial licence on the weights; trails Qwen3-Reranker-4B on MIRACL, RTEB and structured data; about 10 seconds per request on 1,900-token legal documents.
Who is Jina Reranker v3.5 best for?
Elastic users and teams that want a small, long-context reranker on their own hardware.
How much does Jina Reranker v3.5 cost for a team or in production?
$0.05 per million tokens, so 1M queries of about 15,000 tokens each cost around $750 a month. Jina's page does not say whether the query is counted once or per document.
Is Jina Reranker v3.5 safe for company data?
Jina says it does not train on customer inputs and stores them only as needed to serve requests, under Elastic's DPA; no retention period or certification is named. The weights' CC BY-NC 4.0 licence bars commercial self-hosting without Jina On-Prem.
What are the best Jina Reranker v3.5 alternatives?
The closest alternatives we have reviewed are Voyage rerank-3, Cohere Rerank 4 and Qwen3-Embedding and Qwen3-Reranker.

Answers checked Oct 22, 2026.

Make Jina Reranker v3.5? Show this rating on your siteJina Reranker v3.5 rated 4.3 out of 5 on ReviewRamble