Hunyuan-MT is a suite of open-source translation models developed by Tencent, addressing the need for high-quality, multilingual machine translation. It offers both a dedicated translation model (Hunyuan-MT-7B) and an ensemble model (Hunyuan-MT-Chimera) for enhanced accuracy, supporting mutual translation across 33 languages, including several Chinese ethnic minority languages.

How It Works

The project utilizes a comprehensive training framework that progresses from pre-training to continued pre-training, supervised fine-tuning, and reinforcement learning for both translation and ensembling. This multi-stage approach, particularly the novel inclusion of reinforcement learning for ensembling, allows the models to achieve state-of-the-art results for their size. The Chimera model specifically integrates outputs from multiple translation runs to produce a refined, higher-quality translation.

Quick Start & Requirements

• Installation: pip install transformers==4.56.0
• Prerequisites: Python 3.x. For fp8 models, compressed-tensors needs to be upgraded to compressed-tensors-0.11.0, and ignored_layers in config.json must be renamed to ignore.
• Resources: Requires significant computational resources for loading and running the 7B parameter models, with GPU acceleration recommended.
• Links: Hugging Face, ModelScope, Official Website, Demo.

Highlighted Details

• Achieved first place in 30 out of 31 language categories in the WMT25 competition.
• Hunyuan-MT-Chimera-7B is noted as the industry's first open-source translation ensemble model.
• Supports quantization (FP8, INT4) for improved inference efficiency.
• Provides deployment options via TensorRT-LLM, vLLM, and SGLang, with pre-built Docker images available.

Maintenance & Community

The project is actively maintained by Tencent's open-source team. Contact is available via email at hunyuan_opensource@tencent.com.

Licensing & Compatibility

The repository's license is not explicitly stated in the provided README, but it is available on Hugging Face and ModelScope, suggesting a permissive license suitable for most uses. However, users should verify the specific license terms.

Limitations & Caveats

The README mentions a requirement to modify configuration files for fp8 model loading with the transformers library, indicating potential setup friction. The companion branch for the transformers library is noted as pending review, which might imply ongoing development or potential instability.