What is Hunyuan-MT-7B
Hunyuan-MT-7B is a lightweight translation model released by the Tencent Hunyuan team, with only 7 billion parameters, supporting mutual translation among 33 languages and 5 ethnic minority languages/dialects. It won 30 out of 31 first places in the 31 language competitions at the Association for Computational Linguistics (ACL) WMT2025, demonstrating outstanding performance. The model can accurately understand internet slang, ancient poetry, social conversations, etc., and perform free translation based on context. It proposes a training paradigm covering the entire chain from pre-training to ensemble reinforcement. It has fast inference speed, and after processing with Tencent's self-developed AngelSlim compression tool, performance is further improved by 30%. It can be deployed in diverse hardware environments at low cost.
Main Features of Hunyuan-MT-7B
-
Multilingual Translation: Supports mutual translation among 33 languages and 5 ethnic minority languages/dialects, covering Cantonese, Uyghur, Tibetan, Kazakh, Mongolian, etc., meeting translation needs in different language environments.
-
Accurate Context Understanding: Can accurately understand language expressions in special contexts such as internet slang, gaming slang, and ancient poetry, and perform accurate free translation based on context, providing translation results that better fit the context.
-
Efficient Translation Capability: Won 30 out of 31 first places in the 31 language competitions at the Association for Computational Linguistics (ACL) WMT2025, demonstrating outstanding translation performance and the ability to complete translation tasks quickly and accurately.
-
Lightweight and Efficient Inference: With only 7 billion parameters and fast inference speed, after FP8 quantization compression using Tencent's self-developed AngelSlim large model compression tool, inference performance is further improved by 30%, and it can handle more translation requests under the same hardware conditions.
-
Broad Deployment and Application: Can run well in diverse hardware environments from high-end servers to edge devices, with relatively lower deployment, operation, and maintenance costs. It has been integrated into multiple businesses such as Tencent Meeting, WeCom, and QQ Browser, helping improve product experience.
Technical Principles of Hunyuan-MT-7B
-
Complete Training Paradigm: Hunyuan-MT-7B proposes a full-chain training paradigm covering pre-training, CPT tuning, supervised fine-tuning, translation reinforcement, and ensemble reinforcement. Through multi-stage training optimization, the model achieves industry-leading translation performance.
-
Data Cleaning and Filtering: Uses tools such as language identification, document deduplication, perplexity filtering, and parallel sentence cleaning to select "clean" parallel sentence pairs from large amounts of data, ensuring the quality of training data and thereby improving the model's translation accuracy and stability.
-
Model Compression Technology: Based on Tencent's self-developed AngelSlim large model compression tool, FP8 quantization compression is applied to Hunyuan-MT-7B, further improving inference performance by 30%. While ensuring translation quality, it improves the model's computational efficiency and deployment friendliness.
Project Address of Hunyuan-MT-7B
-
Official Website：Tencent Hunyuan
-
Github：https://github.com/Tencent-Hunyuan/Hunyuan-MT/
-
HuggingFace：https://huggingface.co/collections/tencent/hunyuan-mt-68b42f76d473f82798882597
Application Scenarios of Hunyuan-MT-7B
-
Education: Helps students and educators overcome language barriers, access global educational resources, and improve international academic exchange and learning efficiency.
-
Foreign Trade and Business: Promotes international trade, helps merchants communicate effectively with global customers, breaks down language barriers, and expands international markets.
-
Cultural Tourism: To facilitate tourists' travel experience in different language environments, provide real-time translation support, and enhance cultural exchange and interaction.
-
Research Collaboration: Help researchers quickly obtain and share international cutting-edge research results, and promote cross-border research collaboration and academic exchange.
-
Online Social Networking: Provide convenience for cross-language communication among social platform users, and enhance communication and understanding between users from different language backgrounds.
-
Enterprise Services: Support cross-language communication within enterprises, improve internal collaboration efficiency of multinational enterprises, and help the development of global business.