Recently, the official Qwen3-Embedding series models were released by the Tongyi Qianwen team. As new members of the Qwen model family, these models are specifically designed for text representation, retrieval, and ranking tasks. The Qwen3-Embedding series is trained based on the Qwen3 base model, fully inheriting its significant advantages in multi-language text understanding capabilities.
This series of models has performed excellently in various benchmark tests, particularly showcasing outstanding performance in text representation and ranking tasks. Testing was conducted using the MTEB (including English version v2, Chinese version v1, multilingual version, and code version) retrieval datasets, with ranking results based on the top-100 vector recall results of Qwen3-Embedding-0.6B. Notably, the Embedding model with 8B parameter scale ranked first on the MTEB multilingual leaderboard, achieving a score of 70.58, surpassing many commercial API services in performance.
The Qwen3-Embedding series provides three model configurations ranging from 0.6B to 8B parameters to meet the performance and efficiency requirements of different scenarios. Developers can flexibly combine representation and ranking modules to extend functionalities. Additionally, the model supports customizable representation dimensions and instruction adaptation optimization, allowing users to adjust representation dimensions according to actual needs and customize instruction templates to enhance performance for specific tasks, languages, or scenarios.
In terms of multi-language support, the Qwen3-Embedding series performs exceptionally well, supporting over 100 languages, covering mainstream natural languages as well as multiple programming languages. It has strong multi-language, cross-language, and code retrieval capabilities. This series of models adopts a dual-tower structure and a single-tower structure design, respectively for Embedding models and Reranker models, maximizing the retention and inheritance of the base model's text understanding capabilities through LoRA fine-tuning.
During training, the Qwen3-Embedding series adopted a multi-stage training paradigm and underwent deep optimization for specific application scenarios. The Embedding model uses a three-stage training architecture, including contrastive learning pretraining with large-scale weakly supervised data, supervised training with high-quality annotated data, and model fusion strategies to effectively balance the model's generalization ability and task adaptability. The Reranker model directly uses high-quality annotated data for supervised training to improve training efficiency.
The newly released Qwen3-Embedding series models are now open-source on Hugging Face, ModelScope, and GitHub platforms. Users can also directly use the latest text vector model services provided by Alibaba Cloud Bailian Platform. The official stated that this is just the beginning, and with continuous optimization of the Qwen base model, they will continue to enhance the training efficiency of text representation and ranking models and plan to expand the multi-modal representation system to build cross-modal semantic understanding capabilities.
ModelScope:
https://modelscope.cn/collections/Qwen3-Embedding-3edc3762d50f48
https://modelscope.cn/collections/Qwen3-Reranker-6316e71b146c4f
Hugging Face:
https://huggingface.co/collections/Qwen/qwen3-embedding-6841b2055b99c44d9a4c371f
https://huggingface.co/collections/Qwen/qwen3-reranker-6841b22d0192d7ade9cdefea
GitHub:
https://github.com/QwenLM/Qwen3-Embedding