Querit-4B

Querit-Reranker-4B

HighLights

Querit-Reranker-4B is a multilingual cross-encoder reranker initialized from Qwen3-Embedding-4B and further trained with a data-centric, reranking-oriented pipeline. Rather than relying on backbone scale alone, the model first learns broad query-document relevance matching from large-scale ranking supervision and then adapts to target ranking distributions through synthetic-query mining with teacher scores as continuous soft labels.

Selected checkpoints from different data mixtures and training runs are further consolidated with spherical linear interpolation (SLERP), yielding a single deployable reranker without runtime ensembling overhead. By jointly encoding each query-document pair, Querit-Reranker-4B captures fine-grained relevance signals for second-stage ranking and achieves strong performance across multilingual and English retrieval benchmarks.

As of August 29, 2026, Querit-Reranker-4B achieves the best average score of 71.09 among publicly available models on the MTEB Multilingual v2 reranking tasks, averaged over six tasks.

Model Description

<!-- Provide a longer summary of what this model is. -->

• Model type: Text Reranking
• Language(s) (NLP): Multilingual (Chinese, English, Spanish, French, German, Russian, Korean, Japanese)
• Training Stage: Pretraining & Post-training
• Number of Total Parameters: 4.02B
• Number of Paramaters (Non-Embedding): 3.63B
• Number of Layers: 36
• Number of Attention Heads: 32
• Context Length: 128k

Citation

If you find Querit-Reranker useful for your research or applications, please cite our paper:

@misc{zhong2026queritrerankertrainingcompactmultilingual,
      title={Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation}, 
      author={Yunfei Zhong and Jun Yang and Wei Huang and Yinqiong Cai and Haosheng Qian and Yixing Fan and Ruqing Zhang and Lixin Su and Daiting Shi and Jiafeng Guo},
      year={2026},
      eprint={2606.19037},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2606.19037}, 
}