Jina Reranker v3.5
Jina's 0.6B listwise reranker that scores a whole candidate list in one pass, with 131K context and a non-commercial license.
In short. Jina Reranker v3.5 re-sorts search results by reading the question and all the candidate passages together in one pass and scoring each of them. It is a small 0.6B model for many languages, structured data such as tables and JSON, and legal and medical text, and it is free for non-commercial use only.
Summary
Jina Reranker v3.5 is a 0.6B listwise reranker, and the card calls it a drop-in upgrade to jina-reranker-v3, which appeared in September 2025. Instead of scoring each passage against the query separately, it puts the query and a whole list of candidates into one context and scores them together, which Jina calls last-but-not-late interaction, so the list has to fit in the context window, which the card gives as up to 131K tokens. It is built on Qwen3-0.6B with 28 layers, using a repeating pattern of three sliding-window layers (window of 1,024) and two global layers, plus a small projector, and it was trained on legal, medical, financial, multilingual and structured data with a three-stage self-distillation recipe. The Hugging Face repository first appeared on 14 July 2026, the technical report is arXiv 2607.18152, dated 20 July, and Elastic published Jina's announcement on 27 July.
Jina's table, which reranks the top 100 results from jina-embeddings-v5-text-small, gives v3.5 63.20 on BEIR, 74.11 on MIRACL, 70.95 on RTEB and 48.3 on a structured-data test called Struct-IR, against 62.10, 72.20, 68.01 and 38.7 for v3. In the same table Qwen3-Reranker-4B scores 62.28 on BEIR, so v3.5 beats it there, but the 4B model is ahead on MIRACL (76.56), RTEB (77.68) and Struct-IR (55.6). Jina says v3.5 is 1.22 to 1.56 times faster than v3 on an A100, depending on list length, and its blog says it scores more than 50% higher than v3 on case law retrieval. v3 reported 61.94 on BEIR in its own paper, arXiv 2509.25085. All of these numbers are Jina's own.
The license is CC BY-NC 4.0, and commercial use needs a license from Jina. The BF16 weights are 1.19 GB for both v3.5 and v3, and Jina publishes GGUF files from 0.21 GB up to the 1.20 GB BF16 file, with a 0.40 GB Q4_K_M in between, plus MLX builds. Loading needs trust_remote_code, and model.rerank returns scored documents with an optional top_n. In the 30 days to 11 November 2026, Hugging Face counted about 172,000 downloads of v3.5 and 762,000 of v3.
Good for
- Reranking a short list of candidates in one pass, when the whole list fits in the context window
- Legal, medical and financial text and structured data such as tables and JSON, which the v3.5 update targeted
- Cheap serving, from a 1.19 GB BF16 file or a 0.40 GB Q4_K_M GGUF
- Multilingual search, with MIRACL scores reported for v3.5
Watch out for
- The license is non-commercial, so a product needs a separate license from Jina
- The query and all candidates must fit in one context window, so long lists need to be split
- In Jina's own table Qwen3-Reranker-4B is ahead on MIRACL, RTEB and Struct-IR
- Loading needs trust_remote_code, and all scores are Jina's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Jina Reranker v3.5. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| v3.5 (0.6B) BF16 | 597M | 1.19 GB | 1.37 GB | Laptop or phone |
| v3.5 GGUF Q4_K_M INT4 | 597M | 0.4 GB | 0.46 GB | Laptop or phone |
| v3 (0.6B, previous version) BF16 | 597M | 1.19 GB | 1.37 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.
- Attribution or model name must be shown
See also
Frequently asked questions
- What is Jina Reranker v3.5?
- Jina Reranker v3.5 re-sorts search results by reading the question and all the candidate passages together in one pass and scoring each of them. It is a small 0.6B model for many languages, structured data such as tables and JSON, and legal and medical text, and it is free for non-commercial use only.
- Can I use Jina Reranker v3.5 commercially?
- Not without separate permission. Jina Reranker v3.5 is released under the CC BY-NC, which allows research and personal use only.
- Is Jina Reranker v3.5 open source?
- Jina Reranker v3.5 is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run Jina Reranker v3.5 on my own computer?
- Yes. The smallest variant, v3.5 GGUF Q4_K_M, needs about 0.46 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Jina Reranker v3.5 released?
- Jina Reranker v3.5 was released by Jina AI on July 14, 2026.
- How long a context window does Jina Reranker v3.5 have?
- Jina Reranker v3.5 supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- jina-reranker-v3.5 model card (Hugging Face) huggingface.co
- jina-reranker-v3.5: Hybrid-Attention Listwise Reranking with Self-Distillation for Domain-Robust Retrieval (arXiv 2607.18152) arxiv.org
- What's inside Jina's new 600 million parameter listwise reranker (Elasticsearch Labs, 27 July 2026) elastic.co
- jina-reranker-v3 model card (Hugging Face) huggingface.co
- jina-reranker-v3 technical report (arXiv 2509.25085) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Reranker Alibaba Qwen | 596M – 8.19B | 32K | Apache 2.0 (Permissive) | 1.37 GB Laptop or phone | |
| BGE Reranker v2 m3 BAAI | 568M | 8K | Apache 2.0 (Permissive) | 2.61 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone | |
| Jina Embeddings v5 Omni Jina AI | 990M – 1.63B | 32K | CC BY-NC (Non-commercial) | 2.27 GB Laptop or phone |