Hy-MT2 - Tencent Hunyuan's Open-Source Next-Generation Translation Large Model

What is Hy-MT2

Hy-MT2 is Tencent Hunyuan's open-source next-generationtranslationlarge model, offering three sizes: 1.8B, 7B, and 30B-A3B, supporting translation between 33 languages and 5 minority language/dialect translations. Among them, 7B and 30B-A3B achieve the best results among open-source models in multiple evaluations, surpassing models with dozens of times more parameters; the 1.8B lightweight version, after extreme 1.25-bit quantization, is only 440MB and can be deployed for local inference on mobile phones.

Main Features of Hy-MT2

• Multilingual Translation: Supports translation between 33 languages and 5 minority language/dialect translations, covering major global languages and domestic ethnic minority languages.
• Voice Input Translation: The mini program has a built-in voice input feature, allowing users to complete cross-language translation directly by speaking.
• Custom Translation Style: Supports personalized instructions such as setting terminology, output format, and degree of conciseness, making translation results better meet expectations.
• On-Device Offline Translation: The 1.8B model, after 1.25-bit quantization, is only 440MB and can be downloaded to a phone for local inference in no-network environments.
• Word Analysis and Comparison: Provides word definition analysis and comparison of translation results in different styles, helping users understand differences in word usage.

Technical Principles of Hy-MT2

• Multi-Size Model Architecture: Offers three specifications: 1.8B, 7B, and 30B-A3B, respectively targeting lightweight on-device, efficient cloud, and high-precision translation scenarios.
• AngelSlim Extreme Quantization: Uses 1.25-bit quantization technology to compress the 1.8B model to 440MB, enabling low-resource local deployment on mobile phone chips.
• Multi-Scenario Joint Training: Based on unified training with large-scale multilingual corpora, the model's capabilities are optimized in a balanced way across general translation, professional fields, and real business scenarios.
• Instruction Following Reinforcement: Specifically optimizes instruction comprehension for translation tasks, enabling the model to accurately execute complex user requirements such as style transformation and terminology constraints.

How to Use Hy-MT2

• Online Experience
  • Open the Mini Program: Search for and open the Tencent Hy Translation mini program in WeChat.
  • Enter the Content to Translate: Select the source language and target language, enter text, or click the voice button for voice input.
  • Customize Translation Style: Click personalized settings to customize instructions such as translation style, terminology preferences, or format requirements.
  • View Translation Results: View the translation results, switch between different styles for comparison, or click a word to view detailed analysis.
• Offline Use
  • Download the On-Device Model: Download the 1.8B on-device translation model locally in the mini program settings.
  • Enable Offline Mode: After the model download is complete, you can use offline translation in no-network or weak-network environments.
  • Local Inference Translation: Enter text in offline mode, and the model will complete inference locally on the phone and output the translation result.
• Open-Source Model Deployment
  • Obtain the Open-Source Model: Visit GitHub or HuggingFace to obtain the Hy-MT2 open-source models (1.8B / 7B / 30B-A3B).
  • Choose model size: Choose the corresponding model size based on hardware conditions. The 1.8B quantized version requires only about 440MB of storage space.
  • Configure the inference environment: Configure the inference environment according to the open-source documentation, and it can be deployed and used on a local server or terminal device.

Core advantages of Hy-MT2

• Performance surpasses large models: 7B and 30B-A3B surpass DeepSeek-V4-Pro and Kimi K2.6 in evaluations such as FLORES-200, and 30B-A3B surpasses Gemini 3.1 Pro in real-world scenarios and professional fields.
• Extremely lightweight for on-device deployment: The 1.8B model requires only 440MB of storage after quantization, can run smoothly on mobile phone chips such as Apple, Qualcomm, and MediaTek, and inference speed is 1.5 times faster than the previous generation.
• Leading translation instruction following: It performs excellently on the IFMTBench instruction-following test set and can accurately execute complex instructions such as style transformation and translation with specified terminology.
• Deep optimization for professional fields: Deeply tuned for 8 major professional fields including finance, law, healthcare, and education, with GEMBA scores close to or surpassing mainstream commercial APIs.

Hy-MT2 project addresses

• Project official website：https://aistudio.tencent.com/llm/en?tabIndex=0
• GitHub repository：https://github.com/Tencent-Hunyuan/Hy-MT2
• HuggingFace model library：https://huggingface.co/collections/tencent/hy-mt2

Comparison of Hy-MT2 with similar competitors

Application scenarios of Hy-MT2

• Daily cross-language communication: The model is suitable for real-time translation among 33 languages in scenarios such as travel and social interaction.
• Professional document translation: The model can meet the precise translation needs of 8 major vertical fields such as finance, law, healthcare, and education.
• Offline translation scenarios: In no-network or weak-network environments, local offline translation is achieved through the mobile on-device model.
• Personalized content creation: The model supports custom translation styles and assists with specific expression needs such as self-media and academic writing.
• Business office assistance: Supports voice input and document translation, improving the efficiency of cross-border business communication and meeting minutes.