jina-reranker-v3.5
Jina AI
Jina AI's 0.6B-parameter multilingual listwise document reranker, released 27 July 2026 as a drop-in upgrade to jina-reranker-v3 with an unchanged request schema. It keeps the last-but-not-late (LBNL) listwise interface, ranking a query against a whole candidate list in a single forward pass, and is built on Qwen3-0.6B with 28 layers of hybrid 3L2G attention (sliding-window layers of width 1024 interleaved with global-attention layers and a pinned terminal global layer) plus a lightweight 1024-512-512 MLP projector. It is trained with self-distillation for domain robustness and, per the vendor, runs 1.22x-1.56x faster than v3 on short to long contexts. Weights are CC-BY-NC-4.0.
Strengths
- Drop-in replacement for jina-reranker-v3: identical request schema, so only the model string changes
- Listwise last-but-not-late design ranks a query against the full candidate list in a single forward pass, preserving cross-document interaction
- Hybrid 3L2G attention (window-1024 sliding layers interleaved with global layers) gives 1.22x-1.56x faster listwise inference than v3 on short-to-long contexts per the vendor
- Beats its predecessor on every reported axis: +1.10 BEIR, +2.6% relative MIRACL, +4.3% relative RTEB, +24.8% relative Struct-IR
- Strong on structured-data ranking and legal/domain-specific retrieval, where the vendor reports its largest gains
- Can return document embeddings alongside rankings (return_embeddings), and ships GGUF quantizations and MLX builds plus AWS, Azure, GCP and air-gapped deployment options
Caveats
- Weights are CC-BY-NC-4.0 - non-commercial only; the vendor's own card says commercial use requires contacting Jina AI
- It does not win everywhere: in the vendor's own table Qwen3-Reranker-4B scores higher on MIRACL (76.56 vs 74.11), RTEB (77.68 vs 70.95) and Struct-IR (55.6 vs 48.3) - the BEIR headline is the one column it leads
- All reported numbers are vendor-run under a protocol that uses Jina's own jina-embeddings-v5-text-small as the first-stage retriever, which favours the vendor's own stack
- The benchmark protocols are non-standard in two places the vendor flags itself: Struct-IR uses a controlled pool of gold documents plus hardest distractors, and the RTEB figure excludes the MIRACL average
- Context is advertised only as "131K tokens" - a rounded label that does not expand to an unambiguous integer, so no context_window is recorded here
- Its own diagnostics show a low absolute score band (recommended balanced cutoff 0.070, FPR-0.1 cutoff 0.107) and a correct match wins rank 1 in only 66% of the 119 measured queries
- Vendor
- Jina AI
- License
- CC-BY-NC-4.0
- Release Date
- 2026-07-27
- Modalities
- text
Capabilities
Resources
Pricing
Weights are CC-BY-NC-4.0; the model card states "For commercial usage inquiries, feel free to contact us." Hosted access is the Jina Reranker API (https://api.jina.ai/v1/rerank) on prepaid token top-ups; Jina's rate-limit table lists the Reranker API as blocked without a key, 100 RPM & 100,000 TPM on a free key, 500 RPM & 2,000,000 TPM on a paid key and 5,000 RPM & 50,000,000 TPM on a premium key, with tokens counted on the input request. Also available via Elastic Inference Service, AWS SageMaker, Microsoft Azure, Google Cloud, Hugging Face and air-gapped deployment. No per-token price is published as page text, so none is asserted here.