Tencent Hunyuan Releases the New Open-Source Translation Model Hy-MT2, Launches the Mini Program “Tencent Hy Translation”
The biggest improvement is in instruction-following ability
On May 21, Tencent Hunyuan announced the open-source release of its new translation model, Hy-MT2, and launched the translation mini program “Tencent Hy Translation.”
Hy-MT2 is a multilingual model that supports translation among 33 languages. Among them, the 7B and 30B-A3B models achieve state-of-the-art results among open-source models across a variety of translation tasks, outperforming models with parameter scales dozens of times larger. The lightweight 1.8B model also surpasses mainstream commercial APIs such as Microsoft’s. Thanks to AngelSlim 1.25-bit extreme quantization, the model requires only 440MB of storage and can be easily deployed on mainstream mobile chipsets for local inference. Compared with Hy-MT1.5, inference speed has improved by 1.5x.
Hy-MT2 includes three model sizes: Hy-MT2-1.8B, Hy-MT2-7B, and Hy-MT2-30B-A3B, focusing respectively on lightweight on-device deployment, balanced performance, and professional-grade results.
The “Tencent Hy Translation” mini program is built on Hy-MT2. Compared with other translation tools, it not only supports voice input, but also improves the ability to handle custom translation styles and instructions, making translation results closer to expectations and more practical. At the same time, users can experience the high-speed version of the Hunyuan translation model in online environments, and can also download on-device translation models in advance to use offline translation in no-network or weak-network scenarios, solving network constraints in some use cases.
In the general translation benchmark, the three models in the Hy-MT2 series are already very close to the current industry-leading translation model (Gemini 3.1 Pro) in average performance on FLORES-200. Meanwhile, the measured scores of Hy-MT2-7B and Hy-MT2-30B-A3B have already surpassed major domestic general-purpose large models; in a head-to-head comparison of lightweight models, Hy-MT2-1.8B also performs better overall than leading commercial translation APIs.
While maintaining general translation capability, Hy-MT2 has been further optimized for real-world business scenarios and domain-specific translation.
On real-world test sets, Hy-MT2-30B-A3B already outperforms Gemini 3.1 Pro. In particular, on vertical-domain test sets, Hy-MT2-30B-A3B has partially surpassed mainstream translation models in finance, politics, and education.
Compared with the previous version, the biggest improvement in Hy-MT2 is its instruction-following ability. The model can more accurately understand and execute user-specific requirements regarding terminology, style, and output format. Tencent Hunyuan’s in-house dataset IFMT Bench shows that the translation performance of Hy-MT2-7B and Hy-MT2-30B-A3B already exceeds comparable open-source models of similar size and is close to Gemini 3.1 Pro. This benchmark is now also open source.
See the examples below for instruction-following ability. With the personalized setting “Make the translation concise and polished, remove redundant phrasing, and keep each sentence under 15 words,” the model can follow the instruction well and make the translation output better match the requirement.
The upgraded Hy-MT2 model further explores ultra-low-bit quantization schemes. In addition to 4-bit, 8-bit, and FP16 versions, Hy-MT2 also offers 1.25-bit and 2-bit versions based on Tencent Hunyuan’s proprietary technology to meet deployment needs in different hardware environments. The 1.25-bit ultra-low-bit quantized version, implemented with Hunyuan’s proprietary Sherry framework, achieves 1.5x faster inference on Apple A15 compared with Hy-MT1.5’s 4-bit quantized version, further improving practical usability.
To make it easier for developers to use, the open-source Hy-MT2 models have already been released on open-source communities such as Github and Huggingface, with deployment support on multiple platforms including ARM, Qualcomm, Intel, Metax, and Tianshu Zhixin.
Overall, Hy-MT2 is a family of high-quality, highly efficient, multilingual translation models designed for real-world application scenarios, and it shows strong competitiveness in general translation, professional-domain translation, real business scenarios, and translation instruction-following tasks.
Tencent Hunyuan translation models continue to gather real feedback from the community and real application scenarios, constantly improving model capabilities. At the same time, Tencent Hunyuan also hopes to give back to the community through open source and community activities. Tencent Hunyuan is now also working with the official WMT26 “video subtitle translation competition” (https://www2.statmt.org/wmt26/video-subtitle-translation.html). Using Hy-MT series models to participate in the “general machine translation competition” (https://www2.statmt.org/wmt26/translation-task.html) and the “video subtitle translation competition” gives you a chance to win special Hunyuan awards. We warmly invite everyone to participate and jointly promote the development of cutting-edge machine translation technology.
Open-source and demo links: