Tencent releases R3-embedding-0.6B for query-conditional agent skill retrieval

Tencent has released the R3-embedding-0.6B model, a bi-encoder retriever fine-tuned from Qwen3-Embedding-0.6B specifically for retrieving agent skills based on user queries.

• The model embeds queries and skills independently to rank candidates via cosine similarity, serving as the recall stage in a two-stage retrieval system paired with R3-Rerank-0.6B.
• It is designed for query-conditional agent skill routing, addressing the challenge that skills differ from standard documents.
• The release includes code compatible with sentence_transformers and references the paper "Skill Is Not Document: A Query-Conditional Benchmark and Two-Stage Retriever for LLM Agent Skill Routing."

This model enables more accurate matching of user requests to specific agent capabilities by treating skill retrieval as a distinct task from document search.