Tencent has released the R3-rerank-0.6b model, a cross-encoder reranker fine-tuned from Qwen3-Reranker-0.6B. It is designed for query-conditional agent skill retrieval within the R3-Skill framework.
The model scores each (query, skill) pair jointly to repair compatibility judgments that bi-encoders cannot make. It serves as the rerank stage of a two-stage retriever paired with R3-Embedding-0.6B. The release includes code and links to the associated paper on agent skill routing.
This tool enables more accurate matching between user queries and specific LLM agent capabilities.