jina-reranker-v3.5

jina-reranker-v3.5 is a 0.6B parameter multilingual listwise document reranker and a drop-in upgrade to jina-reranker-v3. It keeps the same last-but-not-late (LBNL) listwise interface — ranking a query against many candidates in a single forward pass — while improving domain robustness, structured-data ranking, multilingual retrieval, and inference efficiency. Built on Qwen3-0.6B with 28 layers using hybrid 3L2G attention (sliding-window layers, window 1024, interleaved with global-attention layers and a pinned terminal global layer required by LBNL) and a lightweight MLP projector (1024→512→512), it supports up to 131K tokens of context and ranks documents jointly via causal self-attention and cosine scoring. It reaches 63.20 nDCG-10 on BEIR — above the 4B Qwen3-Reranker at roughly 7× fewer parameters — while running 1.22×–1.56× faster than v3 across short-to-long contexts.

jina-reranker-v3.5 keeps the last-but-not-late (LBNL) listwise design of jina-reranker-v3 — query and candidates share a single context window, and a contextual embedding is read from the last token of each document — but replaces uniform global attention with a hybrid 3L2G scheme: within each group of layers, sliding-window attention layers (window 1024) are interleaved with global-attention layers, and a terminal global layer is pinned as required by LBNL. This lowers attention cost on long candidate lists while preserving cross-document interaction. The model is trained with self-distillation for domain-robust retrieval, and its MLP projector head outputs 512-dimensional embeddings (1024→512→512).

Under a unified top-100 protocol with jina-embeddings-v5-text-small as the first stage, jina-reranker-v3.5 scores 63.20 nDCG-10 on BEIR, 74.11 on MIRACL, 70.95 on RTEB, and 48.3 on Struct-IR, improving over jina-reranker-v3 (62.10 / 72.20 / 68.01 / 38.7) on every axis. Its BEIR score edges out the 4B Qwen3-Reranker (62.28) at roughly 7× fewer parameters and leads all 0.6B-class rerankers evaluated. Relative gains over v3 include +1.10 BEIR, +2.6% MIRACL, +4.3% RTEB (with large lifts on legal retrieval), and +24.8% on structured-data ranking (Struct-IR). Listwise inference runs 1.22×–1.56× faster than v3 from short to long contexts (A100, FlashAttention-2).

jina-reranker-v3.5 is a drop-in replacement for jina-reranker-v3: the request schema is unchanged, so switching the model string from jina-reranker-v3 to jina-reranker-v3.5 is all that is required. It ranks a query against the full candidate list in a single call — pass documents through the standard rerank endpoint and use top_n to limit the returned results, or set return_embeddings to also retrieve document embeddings. The model is especially strong on structured-data ranking, legal and other domain-specific retrieval, and multilingual collections, and its hybrid-attention design keeps latency low on long candidate lists. Available under CC-BY-NC-4.0 on Hugging Face; contact us for commercial use.