jina-reranker-v1-turbo-en

jina-reranker-v1-turbo-en is a 37.8M-parameter English cross-encoder reranker that delivers 95% of the base model's accuracy while processing documents three times faster and using 75% less memory. It was designed for production search systems where the quality-latency tradeoff of full cross-encoders was impractical, offering a practical path to search refinement at scale.

The model compresses the base reranker's 12-layer BERT architecture into a six-layer design with 37.8M parameters (vs. 137M for base). It maintains the core BERT-based cross-attention mechanism for token-level interactions between query and document, but optimizes for speed through reduced layer count and efficient parameter allocation. Training uses knowledge distillation: the larger base model acts as a teacher, guiding the turbo variant to match its ranking behavior with fewer parameters. The model supports sequences up to 8,192 tokens through ALiBi positional encodings, enabling comprehensive document analysis while maintaining fast inference.

On BEIR, the turbo variant achieves NDCG@10 of 49.60, retaining 95% of the base model's performance (52.45) while outperforming bge-reranker-base (47.89, 278M parameters). In RAG applications, it maintains an 83.51% hit rate and 0.6498 MRR. The speed advantage is substantial: three times faster than the base model, with throughput scaling nearly linearly with reduced parameter count. Memory requirements drop from 550MB (base) to 150MB (turbo), enabling deployment on smaller instances and significant cost savings in cloud environments. Performance degradation is minimal on most tasks, with slightly lower scores on extremely nuanced ranking scenarios.

Implement a two-stage pipeline: vector search provides initial candidates, then this model re-ranks the top 100–200. The sweet spot for most applications is reranking 100–200 candidates per query, balancing quality and speed. The model is English-only; for multilingual applications, use jina-reranker-v2-base-multilingual or jina-reranker-v3.5. It requires CUDA-capable hardware but runs on smaller instances than the base model (150MB vs. 550MB GPU memory). Available via Jina Reranker API and AWS SageMaker. For new projects, jina-reranker-v3.5 (multilingual, listwise, 131K context) is the recommended choice. When latency is the primary constraint, this model's 3× speed advantage makes it suitable for real-time search applications.