Tencent hy-mt

Tencent hy-mt is a family of models, designed to run on the edge devices, for machine translation purposes. The family has 1.8B and 7B model variants, trained on mt oriented pre-training, supervised fine tuning and rl. then, distilled to get the 1.8B variant. The training utilizes multilingual monolingual high quality corpora and parallel texts.

The HY-MT family reportedly outperforms middle sized open-sourced models, such as Tower-Plus-72B, Qwen3-32B and translation services such as Microsoft Translator or Doubao Translator. Performs %5 worse than the Gemini 3 Pro and 2% worse than DeepSeek V3.2.

The HY-MT family allows users to utilize terminology features, allowing a consistent terminology use in their translations.

The HY-MT also offers a cost efficient quantized solution for text translation purposes.

Github: https://github.com/Tencent-Hunyuan/HY-MT
Tech report: https://github.com/Tencent-Hunyuan/HY-MT/blob/main/HY_MT1_5_Technical_Report.pdf
ModelScope: https://modelscope.cn/models/Tencent-Hunyuan/HY-MT1.5-1.8B