nv-rerankQA-mistral-4B

nv-rerankQA-mistral-4B-v3 is NVIDIA's retrieval reranking model, part of the NeMo Retriever family and served through NVIDIA NIM. It adapts a Mistral 7B foundation model as a cross-encoder: given a query plus candidate passages, it scores each pair's relevance so a RAG pipeline can reorder retrieval results before generation. Independent benchmark write-ups have reported it near the top of reranker accuracy comparisons, and NVIDIA positions it as production-ready with enterprise support.

The honest picture: this is an API and NIM-container product, not a download-and-run checkpoint, and commercial use is governed by NVIDIA's AI Foundation Models Community License rather than a plain open-source license. Latency is higher than small cross-encoders like MiniLM, so it suits quality-sensitive reranking stages rather than ultra-high-QPS paths. The exact initial release date is not well documented; late 2024 is the era-level estimate. For open-weight reranking, BGE and Jina alternatives remain the self-hostable picks.