Qwen-MT Turbo Release Notes
Qwen has introduced qwen-mt-turbo, an updated version of the Qwen-MT translation model. Built upon Qwen3 and trained on trillions of multilingual and translation tokens, the model utilizes reinforcement learning to improve translation accuracy and linguistic fluency.
Key Capabilities and Features
Qwen-MT provides a balance of high-performance translation and operational efficiency through the following features:
- Broad Language Support: The model supports 92 major official languages and prominent dialects, covering over 95% of the global population.
- Customization Tools: Qwen-MT supports terminology intervention, domain prompts, and translation memory. This allows users to predefine key terminology pairs or provide natural language descriptions of the desired style and domain (e.g., legal vs. social media) to optimize output for mission-critical or domain-specific scenarios.
- Efficient Architecture: Using a lightweight Mixture of Experts (MoE) architecture, the model reduces latency and API costs, with pricing as low as $0.5 per million output tokens.
Translation Quality and Benchmarks
Automatic Evaluation
Qwen-MT was tested on the WMT24 multilingual translation benchmark and multi-domain benchmarks for Chinese-English and English-German translation. Results indicate that Qwen-MT outperforms comparably-sized models, including Gemini-2.5-Flash, GPT-4.1-mini, and Qwen3-8B. The model also remains competitive in quality when compared to larger state-of-the-art models such as Qwen3-235B-A22B, Gemini-2.5-Pro, and GPT-4.1, while offering faster processing speeds.
Human Evaluation
To address the limitations of automatic metrics, Qwen conducted human evaluations using real-world data across ten major languages: Arabic, Chinese, English, French, Italian, Japanese, Korean, Russian, Spanish, and Thai. Three independent professional translators scored samples and used cross-validation to ensure reliability. Qwen-MT demonstrated superior performance in both excellence rates and acceptance rates.
Supported Language Families
Qwen-MT supports a wide array of languages categorized by family:
| Language Family | Supported Languages |
|---|---|
| Indo-European | Afrikaans, Armenian, Assamese, Asturian, Belarusian, Bengali, Bosnian, Bulgarian, Catalan, Croatian, Czech, Danish, Dutch, English, French, Galician, German, Greek, Gujarati, Hindi, Icelandic, Italian, Latvian, Lithuanian, Luxembourgish, Macedonian, Maithili, Marathi, Nepali, Norwegian Bokmål, Norwegian Nynorsk, Occitan, Odia, Polish, Portuguese, Romanian, Russian, Serbian, Sicilian, Sindhi, Sinhala, Slovak, Slovenian, Spanish, Swedish, Tosk Albanian, Ukrainian, Urdu, Venetian, Welsh, Western Persian |
| Sino-Tibetan | Chinese (Cantonese, Simplified, and Traditional), Burmese |
| Afro-Asiatic | Arabic (Standard, Egyptian, Mesopotamian, Moroccan, Najdi, North Levantine, South Levantine, Ta’izzi-Adeni, and Tunisian), Hebrew, Maltese |
| Austronesian | Cebuano, Indonesian, Javanese, Malay, Pangasinan, Tagalog, Waray |
| Dravidian | Kannada, Tamil, Telugu |
| Turkic | Kazakh, North Azerbaijani, Northern Uzbek, Turkish |
| Tai-Kadai | Thai, Lao |
| Uralic | Estonian, Finnish, Hungarian |
| Austroasiatic | Khmer, Vietnamese |
| Other | Basque, Georgian, Japanese, Korean, Swahili |
Implementation and API Usage
Qwen-MT is accessible via the Qwen API. The model accepts translation_options in the extra_body parameter of the API call to handle specific translation needs:
- Terminology Intervention: Users can pass a
termslist containing source and target pairs (e.g., mapping "脑䅈搗大吧" to "biological sensor") to ensure consistent lexicon application. - Domain Prompting: Users can specify the
domainsparameter to guide the model's style. For example, providing a description that the text is from the "Ali Cloud IT domain" ensures the use of professional troubleshooting terminologies and appropriate sentence patterns.