Jina AI Launches v3.5 Reranker: 0.6B Parameters Match 4B Performance
JinaAI_ · x · 2026-08-03
Jina AI has released jina-reranker-v3.5, a new listwise reranker model. Despite having only 0.6B parameters, it achieves an nDCG@10 score of 63.20 on the BEIR benchmark, matching the performance of 4B-parameter models.
Key Highlights:
- Architecture: Utilizes a hybrid attention mechanism (three sliding-window layers + two global layers) to maintain cross-document comparison capabilities while significantly improving efficiency. It reranks up to 1.56x faster than v3.
- Training: Trained on a multi-domain mixture covering legal, medical, financial, multilingual, and structured retrieval. It also employs a three-stage self-distillation recipe.
- Performance: Shows its largest gains in semi-structured retrieval, lifting nDCG@10 by 9.6 points.
The model is open-sourced on Hugging Face, accompanied by a detailed Arxiv paper.
Related event: Jina AI Launches jina-reranker-v3.5(2 posts)→
More from Models
- User extracts internal system prompt of Google Gemini — Empty_Paramedic_5957 · 2026-08-03
- Kimi K3 Tops pmpp-hard Eval: Open Model Achieves SOTA in Hardcore Agent Tasks — xeophon · 2026-08-03
- Diffusion LLMs show more stability under extreme quantization compared to AR models — wFXx · 2026-08-03
- ChatGPT Outperforms Claude in Japanese Voice Input Accuracy — athyuttamre · 2026-08-03
- MTEB Leaderboard Adds Openness Metric; LightOn Scores 100% — antoine_chaffin · 2026-08-03
- Jina AI Launches jina-reranker-v3.5: 0.6B Model Beats Qwen3-4B — JinaAI_ · 2026-08-03