Ali Baba announced a series of three new vector models, Qwen3-Embedding, based on a thousand questions, which optimized training on core tasks such as text representation, retrieval and sequencing, increasing performance by about 40 per cent over the previous generation.
In many authoritative lists such as MTEB (Massive Text Embeding Benchmark), Qwen3-Embedding goes beyond top-of-the-top models of Google, OpenAI, Microsoft, and achieves SOTA (State-of-the-Art, Best Performance) performance of the same model, becoming a global lead in the field of multilingual text embedded.
The Qwen3-Embedding series includes models of the sizes of the 0.6B, 4B and 8B parameters, which are tailored to the different needs and application scenarios. Support for 119 languages, covering many languages, including English, Chinese and Thai, and suitable for global search and sequencing.
The 8B model scored 70.58 (as of 5 June 2025) on the MTEB Multilingual List, ranked first, going beyond Google and OpenAI’s similar models. The accompanying Qwen3-Reranker series (0.6B, 4B, 8B) performed well in the sorting missions, and the 0.6B model was sufficient to cope with most of the scenes, while the 8B model provided a higher level of precision and suited to high-swallow volume demand.
Qwen3The model has been fully open on Hugging Face, GitHub, ModelScope and AliunAPI, supporting the integration of Setence Transformers and facilitating seamless use by developers in search and RAG scenarios.