Alibaba Released Qwen3 Embedding Series for Text Embedding and Ranking

Alibaba has released the Qwen3 Embedding series — new proprietary models of the Qwen family, designed for text embedding, search, and ranking. The models are based on the foundational Qwen3 model, which provides them with multilingual capabilities. The series includes three sizes of embedding models (0.6B, 4B, 8B) and three ranking models (0.6B, 4B, 8B). The embedding models support custom vector dimensions (Matryoshka Representation Learning, MRL) and instructions. The models are open-sourced under the Apache 2.0 license on Hugging Face and ModelScope; the technical report and code are published on GitHub. The architecture of the embedding models is dual-encoder, and that of the ranking models is cross-encoder. Training was conducted in several stages: for the embedding models — contrastive pre-training on weakly labeled data, followed by training on high-quality labeled data and model merging; for the ranking models — direct training on labeled data. The generative capability of Qwen3 was used to generate weakly labeled pairs during the pre-training phase. In the future, Alibaba plans to improve training and deployment efficiency, as well as extend the family to multimodal representations.