Recently, Tencent Hunyuan officially announced the open-source release of its new multilingual translation model Hy-MT2, and simultaneously launched the "Tencent Hy Translation" mini program based on this model, aiming to provide users and developers with high-quality and efficient translation solutions.

Hy-MT2 is a family of models with three sizes, all supporting bidirectional translation across 33 languages, and including translation support for five ethnic languages/dialects. Among them, Hy-MT2-1.8B is a lightweight model optimized for mobile devices. Thanks to Tencent's self-developed AngelSlim 1.25-bit extreme quantization technology, the model size has been compressed to about 440MB, which can be easily deployed on mainstream mobile chipsets from Apple, Qualcomm, and MediaTek for local offline inference. The inference speed has increased by 1.5 times compared to the previous generation, solving the translation needs in weak or no network environments.

QQ20260521-173243.jpg

The other flagship model, Hy-MT2-30B-A3B, introduces a mixture-of-experts architecture for the first time, expanding the total parameter scale while controlling the number of activated parameters per inference, aiming to achieve optimal performance in professional scenarios. In addition, Hy-MT2-7B is positioned as the "sweet spot" size that balances speed and performance.

In terms of performance, the Hy-MT2 series has achieved significant improvements. According to the official evaluation data, the three models reached 88.1%, 96.9%, and 98.1% levels of the current industry-leading closed-source model Gemini 3.1 Pro on the general translation capability (FLORES-200 evaluation set), which is very close.

In professional fields and real business scenarios, the performance is even more prominent. On the DomainMTBench test set covering eight areas such as finance, law, and education, the model scores are close to or reach the level of Gemini 3.1 Pro. Especially in vertical fields such as finance, politics, and education, Hy-MT2-30B-A3B has partially surpassed mainstream translation models. Compared to the previous generation, Hy-MT2 has made significant progress in following translation instructions, better understanding and executing personalized user requirements in terms of terminology, style, and output format.

QQ20260521-173303.jpg

To allow users to experience intuitively, the "Tencent Hy Translation" mini program built on the Hy-MT2 model has already gone live, supporting voice input and custom translation styles. Its iOS and Android applications will also be released soon, supporting offline translation with local model deployment. For developers and researchers, the entire source code of the model series and different quantized versions have been released on mainstream open-source platforms such as Hugging Face, ModelScope, and GitHub, supporting deployment on various hardware architectures.

The release of Hy-MT2 marks a solid step forward for Tencent Hunyuan in building high-quality, efficient, and multilingual translation models for real-world application scenarios. Through the combination of open-source and productization, it promotes the practical application and development of machine translation technology.