Tencent Mix-Huan open-source translation model 1.5 supports deployment on the client side and offline real-time translation, surpassing commercial APIs in effectiveness.

On December 30th, Tencent Hy-MT officially announced the release of the open source translation model version 1.5, which includes two models: Tencent-HY-MT1.5-1.8B and Tencent-HY-MT1.5-7B. They support translation between 33 languages as well as 5 types of Chinese dialects. In addition to common languages like Chinese, English, and Japanese, they also include smaller languages such as Czech, Marathi, Estonian, and Icelandic. Both models are now available on the Tencent Hy-MT official website and can also be downloaded directly from open source communities like Github and Huggingface. HY-MT1.5-1.8B is mainly designed for consumer devices like mobile phones. After quantization, it supports direct deployment on the client side and offline real-time translation, requiring only 1GB of memory to run smoothly. With a parameter size of only 1.8B after quantization, HY-MT1.5-1.8B has shown superior performance compared to similarly sized closed source models in various mainstream translation test sets. It also demonstrates extreme efficiency and cost-effectiveness, with faster inference speed compared to mainstream commercial translation model APIs. The HY-MT1.5-7B model has significantly improved performance compared to the previous version. It is an upgraded version of the model that won the first place in the translation championship of 30 languages in the WMT25 competition. It focuses on improving translation accuracy and reducing the inclusion of comments and mixed languages in the translated text, further enhancing its practicality. In some user scenarios, using both the 1.8B and 7B size models of Hy-MT together can achieve the collaborative deployment of client-side and cloud-based models, enhancing the consistency and stability of the models' performance. Tencent's Hy-MT translation model has not only won first place in international machine translation competitions but has also quickly risen to the top of the HuggingFace model trend list within a week of its first open source release. Currently, the Hy-MT translation model has been implemented in various Tencent business scenarios, including Tencent Meeting, Enterprise WeChat, QQ Browser, and customer service translation.