Tencent’s HY-MT1.5: New Multilingual Translation Models

Tencent’s HY-MT1.5 is a new multilingual machine translation model family designed for both mobile and cloud deployment, featuring two models: HY-MT1.5-1.8B and HY-MT1.5-7B. Supporting translations across 33 languages and 5 dialect variations, these models offer advanced capabilities like terminology intervention, context-aware translation, and format-preserving translation. The 1.8B model is optimized for edge devices with low latency, while the 7B model targets high-end deployments with superior quality. Both models are trained using a comprehensive pipeline that includes general and MT-oriented pre-training, supervised fine-tuning, and reinforcement learning, ensuring high-quality translations and efficient performance. This matters because it enhances real-time, high-quality translation capabilities on a wide range of devices, making advanced language processing more accessible and efficient.

Tencent’s release of the HY-MT1.5 multilingual machine translation models marks a significant advancement in the field of language processing. The models, HY-MT1.5-1.8B and HY-MT1.5-7B, are designed to cater to both mobile devices and cloud systems, providing a versatile solution for translation needs. Supporting 33 languages with additional dialect variations, these models are accessible through open weights on platforms like GitHub and Hugging Face, making them available for a wide range of applications. This democratization of advanced translation technology is crucial as it allows developers and researchers to integrate high-quality translation capabilities into their projects without the need for extensive resources.

The HY-MT1.5-7B model is an enhanced version of the WMT25 championship system and is optimized for complex translation scenarios, including mixed languages and terminology intervention. Its ability to handle contextual and formatted translations makes it particularly valuable for industries that require precise and contextually accurate translations, such as legal, medical, and branding sectors. Meanwhile, the HY-MT1.5-1.8B model offers a more compact solution with less than one-third the parameters of its larger counterpart, yet it maintains comparable performance. This model is particularly suited for edge devices, providing real-time translation capabilities with minimal resource requirements, which is a game-changer for mobile applications and low-power devices.

The development of these models follows a comprehensive training pipeline that includes general pre-training, MT-oriented pre-training, supervised fine-tuning, and reinforcement learning. This approach ensures that the models are finely tuned for translation tasks, as opposed to general language models. The inclusion of human evaluation in the training process further enhances the models’ ability to produce translations that are not only accurate but also culturally and idiomatically appropriate. This focus on quality and efficiency makes the HY-MT1.5 models competitive with larger and more resource-intensive systems, offering a balanced solution for various translation needs.

In practical terms, the HY-MT1.5 models provide several features that are essential for production use, such as terminology intervention, context-aware translation, and format-preserving translation. These capabilities ensure that translations are not only accurate but also maintain the intended meaning and structure. Furthermore, the availability of quantized versions of these models allows them to be deployed on consumer hardware with limited resources, expanding their usability across different platforms. This flexibility and accessibility make the HY-MT1.5 models a valuable tool for businesses and developers looking to incorporate advanced translation technology into their products and services, ultimately enhancing global communication and understanding.

Read the original article here