Text reranker built on Qwen and aligned with the open Qwen3-Reranker series: scores up to 500 candidate documents per query, 100+ languages, 4,000 tokens per item. Alibaba's recommended text reranker; gte-rerank was retired in its favor.
NVIDIA cross-encoder reranker (about 1.7B parameters, December 2025) that scores how relevant a document page, given as an image, text or both, is to a text query; the second stage for visual-document RAG alongside llama-nemotron-embed-vl.
Smallest Qwen3 reranker (0.6B, June 2025): scores query-document relevance with 32K-token inputs, optional task instructions and 100+ languages; the fast second-stage option alongside Qwen3 embeddings.
Mid-size 4B Qwen3 reranker for scoring query-document pairs (32K-token inputs, instruction-aware, multilingual and code), with clearly higher ranking accuracy than the 0.6B at moderate extra cost.
Largest and most accurate Qwen3 reranker (8B) for second-stage ranking of retrieved passages, with 32K-token inputs and multilingual and code coverage; Qwen3-VL-Reranker (January 2026) adds image support.
