Qwen-MT: Translation Where Speed Meets Intelligence

Qwen-MT(qwen-mt-turbo) is based on Qwen3, trained on trillions of multilingual and translation tokens, and enhanced with reinforcement learning to strengthen translation accuracy and fluency.

Its core strengths are threefold.

• Support for 92 languages: Targets diverse real-world translation needs, covering major official languages as well as dialects.
• High customizability: Term intervention, domain prompts, and translation memory enable translation tailored to specialized fields and mission-critical environments.
• Low latency and cost efficiency: Leverages a lightweight Mixture of Experts(MoE) architecture for fast responses, while lowering API costs to around $0.5 per million output tokens.

In quality evaluations, performance was verified on Chinese-English, English-German, and the WMT24 multilingual translation benchmark. It outperformed peer models such as GPT-4.1-mini, Gemini-2.5-Flash, and Qwen3-8B, and showed competitive translation quality even compared with higher-tier models such as GPT-4.1, Gemini-2.5-Pro, and Qwen3-235B-A22B.

Results from direct human evaluation are also highlighted. Using real translation data across 10 major languages — Chinese, English, Japanese, Korean, Thai, Arabic, Italian, Russian, Spanish, and French — 3 professional translators evaluated each sample, followed by cross-validation. The results showed strong performance in both acceptance rate and excellence rate.

The usage method is also simple. In the Qwen API's compatible mode, you specify source_lang and target_lang via translation_options, and if needed, you can add term pairs (terms) and domain descriptions (domains) to guide consistent, specialized translation. In other words, it is positioned as a translation API that offers not only general-purpose translation quality but also control over specialized terminology and style adaptation.