Qwen3 Reranker 0.6B

Qwen3 Reranker 0.6B is a compact cross-encoder reranking model from the Qwen family, built on the Qwen3-0.6B-Base backbone as a 28-layer architecture with 596 million parameters. It is purpose-built for text-ranking tasks, taking query-document pairs and producing logit relevance scores that can be converted into probabilities through sigmoid activation. The model supports custom task instructions via its prompts parameter, with a default tailored to web search retrieval scenarios, and is paired with a Qwen3-specific tokenizer that requires recent transformers versions for proper class handling.

Designed for production retrieval pipelines, the reranker handles 32k-token contexts and covers more than 100 languages, making it well suited for multilingual search and RAG systems that need a lightweight second-pass reranker. It is released under the Apache 2.0 license as ungated open weights, deployable through Sentence Transformers CrossEncoder, raw transformers, or vLLM, with Flash Attention 2 support for memory and speed optimization. Within the broader Qwen3 Embedding series, the 0.6B variant offers an accessible entry point alongside larger 4B and 8B siblings, fitting teams that need cost-efficient reranking without sacrificing the instruction-following flexibility of the Qwen3 family.