Hy-MT2 - Tencent's next-generation open-source translation model

What is Hy-MT2?

Hy-MT2 is a new generation of large-scale translation model open-sourced by Tencent Hunyuan, offering three sizes: 1.8B, 7B, and 30B-A3B. It supports mutual translation between 33 languages and translation between 5 minority languages/dialects. Among them, the 7B and 30B-A3B models achieved the best performance among open-source models in multiple evaluations, surpassing models with dozens of times more parameters. The lightweight 1.8B version, after 1.25-bit extreme quantization, is only 440MB and can be deployed for local inference on mobile devices.

Main functions of Hy-MT2

• Multilingual translationIt supports mutual translation between 33 languages and translation between 5 ethnic minority languages/dialects, covering mainstream languages worldwide and ethnic minority languages in China.
• Voice input translationThe mini-program has a built-in voice input function, allowing users to directly translate between languages by speaking.
• Custom translation styleIt supports setting personalized instructions such as terminology, output format, and level of simplification, making the translation results more in line with expectations.
• End-to-end offline translationThe 1.8B model, after 1.25-bit quantization, is only 440MB, which can be downloaded to a mobile phone for local inference in an offline environment.
• Word analysis and comparisonIt provides word definitions and comparisons of translations in different styles to help users understand the differences in word usage.

Hy-MT2 Technical Principles

• Multi-size model architectureIt offers three specifications: 1.8B, 7B, and 30B-A3B, targeting lightweight edge translation, efficient cloud translation, and high-precision translation scenarios, respectively.
• AngelSlim Extreme QuantizationThe 1.8B model is compressed to 440MB using 1.25-bit quantization technology, enabling low-resource local deployment on mobile phone chips.
• Multi-scenario joint trainingBased on unified training of large-scale multilingual corpora, the model's capabilities are optimized in a balanced manner across general translation, professional domains, and real-world business scenarios.
• Instruction compliance enhancementSpecifically optimized for translation tasks, the model can accurately execute complex user requirements such as style transformation and terminology constraints.

How to use Hy-MT2

• Online experience
  • Open the mini programSearch for and open the Tencent Hy Translate mini-program on WeChat.
  • Enter the content to be translatedSelect the source and target languages, enter text, or click the voice button for voice input.
  • Custom translation styleClick on Personalization Settings to customize translation style, terminology preferences, or formatting requirements.
  • View translation resultsView the translation results; you can switch between different styles for comparison or click on a word to view a detailed explanation.
• Offline use
  • Download client-side modelDownload the 1.8B client-side translation model to your local machine in the mini-program settings.
  • Enable offline modeOnce the model has finished downloading, offline translation can be used in environments with no or weak network connectivity.
  • Local reasoning translationWhen you input text offline, the model will perform inference and output the translation result locally on your phone.
• Open source model deployment
  • Obtain open source modelAccess GitHub or HuggingFace to obtain the Hy-MT2 open-source model (1.8B / 7B / 30B-A3B).
  • Select model sizeChoose the appropriate size model based on your hardware requirements; the 1.8B quantized version only requires approximately 440MB of storage space.
  • Configure the inference environmentConfigure the inference environment according to the open-source documentation, and it can be deployed and used on a local server or terminal device.

Hy-MT2's core advantages

• Performance surpasses large modelsThe 7B and 30B-A3B outperformed the DeepSeek-V4-Pro and Kimi K2.6 in reviews such as FLORES-200, and the 30B-A3B surpassed the Gemini 3.1 Pro in real-world scenarios and professional fields.
• Extremely lightweight edge deploymentThe 1.8B model requires only 440MB of storage after quantization and can run smoothly on mobile chips from Apple, Qualcomm, MediaTek, etc., with an inference speed 1.5 times faster than the previous generation.
• Translation instructions follow the leadIt performs excellently on the IFMTBench instruction compliance test set, and can accurately execute complex instructions such as style transformation and specified term translation.
• In-depth optimization of professional fieldsWith in-depth optimization across eight professional fields including finance, law, healthcare, and education, GEMBA scores approach or surpass mainstream business APIs.

Hy-MT2 project address

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

Hy-MT2 Comparison with Similar Products

Hy-MT2 Application Scenarios

• Daily cross-language communicationThe model is applicable to real-time translation between 33 languages in scenarios such as tourism and social interaction.
• Professional document translationThe model can meet the accurate translation needs of eight major vertical fields, including finance, law, healthcare, and education.
• Offline translation scenariosIn environments with no or weak network connectivity, local offline translation is achieved through a mobile-side model.
• Personalized content creationThe model supports custom translation styles to assist in specific expression needs such as social media and academic writing.
• Business Office AssistanceIt supports voice input and document translation, improving the efficiency of cross-border business communication and meeting recording.