Querit-Reranker boosts multilingual search with compact, label-free training

Researchers from multiple institutions present Querit-Reranker, a family of compact multilingual rerankers designed for efficient, label-free distribution adaptation. The models are initialized from two backbones: Querit-Reranker-A0.4B uses an in-house MoE with 0.4B activated parameters, and Querit-Reranker-4B is based on Qwen3-Embedding-4B. The training pipeline first learns general relevance from large-scale ranking data, then adapts to target distributions using synthetic queries generated by a teacher model, whose scores serve as continuous soft labels—eliminating the need for human annotations. To consolidate task-specific strengths, checkpoints are merged via spherical linear interpolation, yielding a single deployable model without ensembling overhead.

Performance benchmarks show significant gains. With Qwen3-Embedding-0.6B as the first-stage retriever, the 0.4B model boosts nDCG@10 from 54.11 to 59.28 on BEIR and from 59.87 to 67.70 on MIRACL, outperforming larger embedding-based baselines on MTEB Multilingual v2 Reranking. The 4B version achieves state-of-the-art among publicly available models. Both models are released on Hugging Face, enabling practitioners to deploy efficient multilingual second-stage reranking without costly annotation efforts.