HY-MT 2

Copyright Notice

© 2026 Tencent-Hunyuan. All intellectual property rights of Tencent Hunyuan HY-MT2 Series Models are owned by Tencent. Tampering with or removing copyright notices is strictly prohibited.

Compliance Notes

• Comply with the Tencent Hunyuan Community License Agreement and relevant laws; use only for legitimate purposes.
• Users in restricted regions (e.g., EU, UK, South Korea) are prohibited from downloading/using the models. Downloading constitutes confirmation that your region complies with the requirements, and you shall bear sole responsibility for any non-compliant use.
• The models shall not be used to improve non-Hunyuan series AI models. You shall bear sole responsibility for any non-compliant use.

Basic Information

Hy-MT2 is a fast-thinking multilingual translation model family developed by Tencent Hunyuan Team. It is designed for complex real-world application scenarios, covering daily communication, business office, professional fields, video subtitles and other full-scenario translation needs. It adopts an advanced large language model architecture at the bottom, integrating core technologies such as MoE mixture-of-experts architecture, extreme quantization compression, instruction-following translation, and structured data translation. It also supports open-source evaluation benchmarks, complete training pipelines and multi-deployment adaptation solutions.

Full List of Supported Languages

Chinese, English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, Malay, Indonesian, Filipino, Hindi, Traditional Chinese, Polish, Czech, Dutch, Khmer, Burmese, Persian, Gujarati, Urdu, Telugu, Marathi, Hebrew, Bengali, Tamil, Ukrainian, Tibetan, Kazakh, Mongolian, Uyghur, Cantonese.

Core Functions and Product Features

1. Full coverage of model specifications: Three sizes are available: 1.8B lightweight version, 7B standard version, and 30B-A3B MoE version, adapting to terminal local deployment, cloud service, high-performance inference and other demands.
2. Powerful cross-language translation: Supports mutual translation of 33 languages in full, accurately follows multilingual translation instructions, and adapts to various translation tasks such as general text, business copy and professional fields.
3. Extreme lightweight deployment: The self-developed AngelSlim 1.25-bit extreme quantization technology compresses the 1.8B model to only 440MB storage, increases inference speed by 1.5x, and realizes offline terminal deployment easily.
4. Top-tier translation performance: The 7B and 30B-A3B versions outperform open-source models like DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode; the 1.8B lightweight version surpasses mainstream commercial translation APIs such as Microsoft and Doubao overall.
5. Diversified instruction translation: Supports 7 major instruction translation scenarios: default translation, terminology fixation, style customization, personalization preference, delimiter retention, structured data translation and context-based translation.
6. Complete open-source support: Open-sources the self-developed IFMTBench benchmark for translation instruction-following evaluation; compatible with mainstream deployment frameworks including transformers, vLLM, SGLang and llama.cpp; provides full-parameter fine-tuning and LoRA fine-tuning pipelines.
7. Event ecological cooperation: Officially partners with WMT26, offering special awards for participants in general machine translation and video subtitle translation tasks to promote the development of machine translation technology.

New Upgrades Compared with the Old Version

1. Added the 30B-A3B MoE mixture-of-experts large model, greatly improving the accuracy of complex long texts and professional domain translation.
2. Expanded instruction translation templates, adding advanced capabilities such as structured data translation, delimiter retention and context linkage translation.
3. Special optimization of inference parameters, with exclusive generation parameters for small and large models, significantly improving translation fluency and accuracy.
4. Integrated with ClawHub and SkillHub platforms, enabling rapid integration into various applications for translation capability implementation.