Revolutionary Multilingual Processing Capabilities and Globalized Applications
Qwen3's language support reaches an industry-leading 119 languages and dialects, covering a global Internet user base of 98%. Its language matrix not only contains mainstream languages (such as English-Chinese-French-German), but also a large number of scarce resource languages (such as Bashkir, Papiamento, etc.). The technical documentation shows that the capability originates from 36 trillion token of super large-scale pre-training data, of which the proportion of non-English data reaches 45%, much higher than the industry average of 20-30%.
For the implementation mechanism, the team adopts a triple innovation: multimodal data cleaning based on Qwen2.5-VL, language-specific embedding space optimization, and dynamic vocabulary expansion techniques. Particularly in dialect processing (e.g., seven dialect variants of Arabic), the model achieves dialect intercomprehension through phoneme-level representation learning. Test data show that Qwen3's translation quality for small languages is 15 percentage points higher than GPT-4 on the FLORES-200 benchmark.
This feature brings breakthroughs in cross-border commerce, multilingual content creation and other scenarios, such as automatically generating marketing copy that conforms to regional cultural habits. It is reported that the model has been pilot applied in the United Nations multilingual document processing system, with an accuracy rate of 92%.
This answer comes from the articleQwen3 Released: A New Generation of Big Language Models for Thinking Deeply and Responding FastThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil