Jev 1.13 beats Cohere, Voyage in latency

Fast structured-decision models are threatening to disrupt the dedicated cross-encoder reranking market sooner than expected.

• –Jev 1.13 edges out established reranking offerings on both median speed (149 ms) and tail latency (205 ms p95), which is crucial for real-time RAG stacks where latency budgets are tight.
• –Jev demonstrates much tighter tail variance compared to Cohere 4 Fast, which experienced severe tail latency spikes reaching 824 ms at p95.
• –Because Jev avoids token-by-token generation overhead and provides calibrated probability scores, it can natively perform binary relevance checks and document filtering at high concurrency without the cost penalty of standard LLMs.
• –While latency numbers are very promising, comprehensive quality benchmarks evaluating nDCG@10 and MRR on datasets like BEIR are still needed to verify whether its ranking quality matches or exceeds top specialized cross-encoders.