jina-reranker-v1-tiny-en

jina-reranker-v1-tiny-en is a 33M-parameter English cross-encoder reranker — the smallest in Jina's v1 family. Through aggressive knowledge distillation, it retains 92.5% of the base model's accuracy while processing documents five times faster and using 13% less memory than the turbo variant. It is designed for edge computing, mobile applications, and high-throughput search systems with strict latency budgets.

The model employs a streamlined four-layer architecture based on JinaBERT with symmetric bidirectional ALiBi positional encodings. Its development leverages knowledge distillation from the 137M-parameter jina-reranker-v1-base-en, which serves as the teacher model. The student model learns optimal ranking behaviors without requiring extensive real-world training data, matching the teacher's soft ranking distributions while using only 33M parameters. The four-layer design and reduced hidden dimensions enable the 5× speedup over the base model. The model supports 1,024-token context with automatic chunking for longer documents.

On BEIR, the model achieves NDCG@10 of 48.54, retaining 92.5% of the base model's performance (52.45) while being just a quarter of its size. In LlamaIndex RAG benchmarks, it maintains an 83.16% hit rate, nearly matching larger models while processing documents significantly faster. Throughput is nearly five times faster than the base model, with 13% less memory usage than the turbo variant. These metrics rival or exceed much larger models like mxbai-rerank-base-v1 (184M) and bge-reranker-base (278M). The performance-size tradeoff makes it uniquely suited for resource-constrained deployments.

Prioritize scenarios where processing speed and resource efficiency are critical: edge computing, mobile applications, and high-throughput search systems with strict latency budgets. For applications requiring absolute maximum ranking precision, the base model or jina-reranker-v3.5 is preferable. The model requires CUDA-capable GPU for optimal performance but runs on less powerful hardware than larger counterparts. It integrates with major vector databases and RAG frameworks, and is available through the Jina Reranker API and AWS SageMaker. The model is English-only; for multilingual applications, use jina-reranker-v2-base-multilingual or jina-reranker-v3.5. When fine-tuning for specific domains, carefully balance training data quality with the model's compact architecture to maintain performance characteristics.