Qwen3-Reranker

Qwen3-Reranker is Alibaba's reranking model family, released June 5, 2025 alongside Qwen3-Embedding. It comes in 0.6B, 4B, and 8B sizes, supports 100+ languages, and is released under Apache 2.0. As a cross-encoder it scores query-document pairs for relevance, making it the natural second stage after a Qwen3-Embedding retriever in a RAG pipeline.

The 0.6B variant is the practical standout: it is CPU-viable through GGUF builds under llama.cpp, so small deployments get reranking without a GPU. The larger sizes trade latency for ranking quality on harder retrieval tasks.

The honest caveat is sizing discipline: the 8B model is heavy for a reranking stage that runs per query at scale, so most teams should start with the 0.6B or 4B and only scale up if relevance metrics justify it.