Tencent Ships Hy-MT2-30B-A3B, 33-Language Translator That Runs Locally

By vramkickedin | | 2 min read
Sleek polygonal globe placed precisely on the right third of the frame with delicate etched characters from various writing systems.

Tencent has released Hy-MT2-30B-A3B, an open-source multilingual translation model that uses a mixture-of-experts design to deliver fast, high-quality results. It handles translation among 33 languages and can follow complex instructions like style guides, term lists, or structured data formats. The model is part of the Hy-MT2 family, which scales from a tiny 1.8B version to this large 30B-A3B variant.

Tencent’s Hunyuan team that delivered Hy-MT1.5-1.8B-1.25bit and Pixal3D, built Hy-MT2 to address real-world business translation needs where off-the-shelf tools often fail. The 30B-A3B model outperforms other open-source heavyweights like DeepSeek-V4-Pro and Kimi K2.6, and even the smallest 1.8B version beats commercial APIs from Microsoft and Doubao in overall quality. By open-sourcing the weights and code, the team gives users full control to run translations locally without sending data to the cloud.

Fast and flexible translation for 33 languages

Key features
  • Translates between 33 languages including Cantonese and Tibetan.
  • Follows detailed instructions on style, terminology, and formatting.
  • Mixture-of-experts architecture activates only 3B parameters per token.
  • Outperforms open-source rivals DeepSeek-V4-Pro and Kimi K2.6.
  • Deploy locally with Hugging Face transformers, vLLM, or SGLang.
  • Open-source IFMTBench for evaluating translation instruction following.

Privacy-conscious professionals and small agencies can run Hy-MT2-30B-A3B entirely on local hardware, keeping sensitive documents secure. Serious hobbyists with consumer GPUs benefit from the mixture-of-experts design, which saves memory and speeds up inference. Localization teams will appreciate the model’s ability to preserve delimiters, structured data, and personal preferences during translation.

Developer notes and community involvement

The model does not include a default system prompt, and the team recommends specific generation parameters (temperature 0.7, top_p 1.0) for best results. Full-parameter and LoRA fine-tuning are supported, with documentation available for training and deployment. Tencent also partnered with WMT26’s video subtitle translation task, inviting the community to use Hy-MT models.

“Hy-MT2 is a family of fast-thinking multilingual translation models designed for complex real-world scenarios.” — Source: arXiv

Project Links