Tencent-Hunyuan announced the open-sourcing of its Hunyuan-MT-7B and Hunyuan-MT-Chimera. These models are designed for translation tasks, supporting mutual translation among 33 languages, including five ethnic minority languages in China: Tibetan, Kazakh, Mongolian, Uyghur, and Cantonese.

The Hunyuan-MT-7B serves as the base translation model, while the Hunyuan-MT-Chimera is an ensemble model that combines multiple translation outputs for improved results. In the WMT25 competition, the models ranked first in 30 out of 31 participated language categories.

Key details include:

1. Performance: The models achieve leading results for their scale, with a full training framework from pretraining to ensemble reinforcement learning. Additional experimental results are available in the technical report.
2. Model Variants: Available downloads include FP8-quantized versions for both models.
3. Prompt Templates: Specific formats are provided for ZH↔XX translations, general XX↔XX translations, and ensemble use.
4. Usage: Compatible with transformers library (version 4.56.0 recommended). Example code is given for loading and inference, with suggested parameters like top_k=20, top_p=0.6, repetition_penalty=1.05, and temperature=0.7.
5. Supported Languages: Includes Chinese (Simplified and Traditional), English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, Malay, Indonesian, Filipino, Hindi, Polish, Czech, Dutch, Khmer, Burmese, Persian, Gujarati, Urdu, Telugu, Marathi, Hebrew, Bengali, Tamil, and Ukrainian.
6. Fine-Tuning: Guidance provided using LLaMA-Factory, with data in ShareGPT format.
7. Quantization and Deployment: Supports FP8, INT4, and INT8 quantization via AngleSlim. Deployment options include TensorRT-LLM, vLLM (version 0.10.0+), and SGLang, with Docker images and API server examples.

For more information, refer to the GitHub repository at Tencent-Hunyuan/Hunyuan-MT or the technical report.