Alibaba launches Qwen-MT: a machine translation model for 92 languages with advanced customization

Alibaba introduced Qwen-MT (qwen-mt-turbo) via the Qwen API, based on Qwen3 and trained on trillions of multilingual and translation tokens. The model supports 92 languages covering over 95% of the global population. It incorporates reinforcement learning to improve accuracy and fluency. Qwen-MT offers high customizability with features like terminology intervention, domain prompts, and translation memory. Its lightweight Mixture of Experts (MoE) architecture provides low latency and cost efficiency, with API costs as low as $0.5 per million output tokens. In automatic evaluations on Chinese-English, English-German, and WMT24 benchmarks, Qwen-MT outperformed comparably-sized models including GPT-4.1-mini, Gemini-2.5-Flash, and Qwen3-8B, and maintained competitive quality against larger models like GPT-4.1, Gemini-2.5-Pro, and Qwen3-235B-A22B. Human evaluations across ten major languages confirmed superior acceptance and excellence rates. The model is accessible via the Qwen API with support for advanced features such as terminology intervention and domain-specific prompts.