TENCENT(HK:00700) Zheng Quan Shi Bao Wang·2025-12-30 13:43Core Insights - Tencent Mixyuan has launched and open-sourced its translation model 1.5, which includes two models: Tencent-HY-MT1.5-1.8B and Tencent-HY-MT1.5-7B, supporting translation across 33 languages and 5 Chinese dialects [1] Model Performance - The HY-MT1.5-1.8B model is designed for consumer devices, requiring only 1GB of memory for smooth operation, and outperforms most commercial translation APIs in terms of efficiency and cost-effectiveness [1] - The model processes 50 tokens in an average time of 0.18 seconds, significantly faster than other models which take around 0.4 seconds [1] - The HY-MT1.5-7B model has improved translation accuracy and reduces issues with mixed languages and annotations, building on its predecessor that won the WMT25 competition [1] Practical Applications - The two models can be used simultaneously for edge and cloud deployment, enhancing consistency and stability in translation results [2] - The HY-MT1.5-1.8B model achieved approximately 78% in the FLORES-200 quality assessment, showcasing its speed advantage and suitability for real-time translation scenarios [2] - Both models support terminology customization, allowing users to create specialized glossaries for various industries, ensuring accuracy in professional translations [3] Advanced Features - The models possess advanced context understanding capabilities, improving coherence in long dialogues and multi-turn interactions [3] - They maintain formatting during translation, ensuring that the output is both accurate and practical [3] Technical Innovations - The effectiveness of the HY-MT1.5-1.8B model is attributed to the On-Policy Distillation strategy, which allows the smaller model to learn from the larger model [4] - The Mixyuan translation model has achieved significant recognition, including winning first place in international machine translation competitions and topping the Hugging Face model trends shortly after its release [4] - The models are already being applied in various Tencent services, including Tencent Meeting, WeChat Work, and QQ Browser, and are available on open-source platforms like GitHub and Hugging Face [4]