SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Tencent-HY-MT1.5 - Tencent Hunyuan's open-source translation model

Author: AI小集 · Original source: AI工具集

What is Tencent-HY-MT1.5

Tencent-HY-MT1.5 is an open-source translation model from Tencent Hunyuan, comprising two versions: Tencent-HY-MT1.5-1.8B and Tencent-HY-MT1.5-7B. The model supports mutual translation among 33 international languages and translation for 5 ethnic/minority languages/dialects, covering a variety of low-resource languages. The 1.8B version has been quantized and requires only 1GB of memory to achieve on-device offline real-time translation on consumer-grade devices such as mobile phones, with fast processing speed. The 7B version is aimed at high-performance scenarios, with high translation accuracy, effectively reducing the occurrence of annotations and language mixing in translated text.

Tencent-HY-MT1.5

Main features of Tencent-HY-MT1.5

  • Multilingual support: Covers mutual translation among 33 languages as well as 5 ethnic/minority languages/dialects. In addition to common languages such as Chinese, English, and Japanese, it also includes low-resource languages such as Czech, Marathi, Estonian, and Icelandic.
  • On-device deployment capability: HY-MT1.5-1.8B has been quantized and requires only 1GB of memory to achieve offline real-time translation on consumer-grade devices such as mobile phones. The average time to process 50 tokens is only 0.18 seconds, far faster than the 0.4 seconds of mainstream commercial translation APIs.
  • High translation quality: In authoritative test sets such as FLORES-200, HY-MT1.5-1.8B achieves the 90th percentile level of ultra-large closed-source models such as Gemini-3.0-Pro, comprehensively surpassing medium-sized open-source models.
  • Enhanced practicality: Supports custom terminology databases, long-text context understanding, and formatted text translation. Users can import professional terminology glossaries; the model can optimize subsequent translations based on the preceding context and preserve the original layout format.

Technical principles of Tencent-HY-MT1.5

  • On-Policy Distillation strategy: The 1.8B model adopts the On-Policy Distillation (large-size model distillation) strategy, in which the 7B large model serves as a teacher to guide training in real time, avoiding rote memorization of standard answers by the small model. By correcting deviations when predicting sequence distributions, the small model learns from mistakes, significantly improving the small model's generalization ability and translation quality.
  • Custom terminology databases: Users can build dedicated terminology correspondence tables for different industries and professional scenarios. After importing a terminology database, the model will prioritize the user-defined standard terminology during translation, ensuring the accuracy and consistency of key terms.
  • Contextual translation: The model has long-text and dialogue context understanding capabilities and can continuously optimize subsequent translation results based on the preceding context. It is suitable for translating long-form content such as meeting minutes, interview content, novel chapters, and technical documents.
  • Formatted translation: The model can keep format information unchanged before and after translation, making it suitable for translating structured content such as web pages and documents.

How to use Tencent-HY-MT1.5

  • Obtain the model: Visit theTencent Hunyuanofficial website for an online experience; you can visit GitHub or the HuggingFace platform to download the required Tencent-HY-MT1.5-1.8B or Tencent-HY-MT1.5-7B model for deployment and use.
  • On-device deployment: Tencent-HY-MT1.5-1.8B has been quantized and requires only 1GB of memory to enable direct on-device deployment and offline real-time translation on consumer-grade devices such as mobile phones.
  • Cloud-side deployment: Tencent-HY-MT1.5-7B is suitable for high-performance scenarios and can be deployed in the cloud, working in coordination with the 1.8B model to improve the consistency and stability of results.

Project address of Tencent-HY-MT1.5

  • GitHub repository：https://github.com/Tencent-Hunyuan/HY-MT）
  • HuggingFace Model Library：https://huggingface.co/collections/tencent/hy-mt15

Application scenarios of Tencent-HY-MT1.5

  • Mobile device translation: Tencent-HY-MT1.5-1.8B has been quantized and requires only 1GB of memory to enable on-device offline real-time translation on consumer-grade devices such as mobile phones, making it suitable for instant translation needs on mobile devices, such as language communication during travel and cross-language conversations in instant messaging, etc.
  • High-performance translation service: Tencent-HY-MT1.5-7B is aimed at high-performance scenarios and is suitable for deployment in the cloud, providing services for enterprise-level applications that require high translation quality, such as document translation for multinational enterprises and technical document translation in professional fields, etc.
  • Multilingual content creation: It supports mutual translation among 33 international languages and translation between 5 ethnic minority languages/Chinese dialects, and can meet the creation needs of content creators in multilingual environments, such as multilingual video subtitle production and cross-language literary creation, etc.
  • Professional terminology translation: It supports customization of terminology databases. Users can import professional terminology glossaries, and the model will prioritize the standard terms defined by users during translation. It is suitable for translation needs in professional fields, such as translation in the medical, legal, and technology industries, etc.
  • Long text translation: It has the ability to understand long texts and dialogue context, and can continuously optimize subsequent translation results based on the preceding context. It is suitable for translating long-form content such as meeting minutes, interview content, novel chapters, and technical documents.
  • Formatted content translation: It can keep format information unchanged before and after translation, and is suitable for translating structured content such as web pages and documents, ensuring that the translated format is consistent with the original text.