In today's rapidly advancing field of artificial intelligence, Alibaba officially released its brand-new Qwen3-Embedding series of vector models on June 6. Built upon the Qwen3 foundation, this model is optimized for tasks such as text representation, retrieval, and ranking, marking another significant breakthrough for Alibaba in the AI domain.
Compared to its predecessor, Qwen3-Embedding has achieved over a 40% performance improvement in core tasks like text retrieval, clustering, and classification. This advancement not only enhances its competitiveness technologically but also places it at the top of professional rankings, surpassing leading models such as Google's Gemini Embedding, OpenAI's text-embedding-3-large, and Microsoft's multilingual-e5-large-instruct, achieving state-of-the-art (SOTA) performance among similar models.
Figure Source Note: Image generated by AI
In simple terms, a vector model can be seen as an "interpreter" for AI. It converts unstructured information like text and images into vectors that machines can easily understand, enabling efficient classification, retrieval, and sorting of information. Leveraging the Qwen3 model, the Tongyi team successfully developed this new vector model using methods such as contrastive training, SFT, and model fusion. The resulting models include the text embedding model Qwen3-Embedding and the text reranking model Qwen3-Reranker.
Notably, the Qwen3 vector model boasts powerful multi-language capabilities, supporting more than 100 languages, including multiple programming languages. This showcases its excellent ability in multilingual, cross-language, and code retrieval tasks, providing global developers with broader application opportunities.
This open-source release of the Qwen3 vector model includes nine different versions, covering various sizes (0.6B, 4B, 8B, etc.) and GGUF versions. Developers can choose suitable models based on their needs, freely combine modules, or even customize vectors or instructions to achieve deep optimization for specific tasks, languages, and scenarios. This flexibility will significantly enhance developers' efficiency.
Currently, the Qwen3 Embedding and Reranker models have been made available on multiple platforms, including ModelScope, Hugging Face, and GitHub. Developers can also use API services via Alibaba Cloud Bailian. Since its open-source launch on April 29, the Qwen3 large model has achieved top rankings in several international benchmarks.
