Tencent YouTu launches high-performance general-purpose text embedding model Youtu-Embedding

Tencent YouTu Lab has open-sourced Youtu-Embedding, a high-performance general-purpose text embedding model. Designed for enterprise applications, the model can handle six major tasks, including text retrieval and intent understanding, and topped the Chinese text embedding benchmark CMTEB with a score of 77.46. It employs a three-stage training process: LLM-based pre-training → weakly supervised alignment → co-discriminative fine-tuning. Combined with an innovative fine-tuning framework and refined data engineering, it solves the challenges of multi-task learning.