Tencent-Hunyuan's HY-MT1.5 project delivers advanced neural machine translation models, HY-MT1.5-1.8B and HY-MT1.5-7B, supporting 33 languages and specialized variations. It targets developers needing efficient, high-quality translation, offering competitive performance, edge deployment capabilities, and features for complex scenarios.

How It Works

The project features HY-MT1.5-1.8B and HY-MT1.5-7B models, covering 33 languages and 5 ethnic/dialect variations. HY-MT1.5-7B is an enhanced WMT25 championship model, optimized for explanatory and mixed-language translation. The smaller HY-MT1.5-1.8B model matches larger counterparts' performance with fewer parameters, enabling real-time and edge deployment. Both models support terminology intervention, contextual translation, and formatted translation for nuanced requirements.

Quick Start & Requirements

• Installation: pip install transformers==4.56.0. A companion transformers branch may be needed.
• Prerequisites: Python, transformers (v4.56.0 recommended). GPU acceleration is advised. Quantized models require tools like AngelSlim or specific deployment frameworks.
• Deployment: Frameworks like TensorRT-LLM, vLLM, or SGLang are recommended, with pre-built Docker images available.
• Links: Hugging Face (https://huggingface.co/models?search=tencent/HY-MT1.5), ModelScope (https://www.modelscope.cn/models?search=Tencent-Hunyuan), Technical Report (https://arxiv.org/abs/2512.24092).

Highlighted Details

• HY-MT1.5-7B is an upgraded WMT25 championship model.
• HY-MT1.5-1.8B offers industry-leading performance for its size, enabling edge deployment.
• Supports 33 languages plus 5 ethnic/dialect variations.
• Advanced features: terminology intervention, contextual translation, formatted translation.

Maintenance & Community

Tencent-Hunyuan actively maintains the project, releasing models on Hugging Face and ModelScope. Contact for feedback is via hunyuan_opensource@tencent.com.

Licensing & Compatibility

The README does not specify a software license, making terms for use, modification, and distribution undefined. Commercial use compatibility is uncertain.

Limitations & Caveats

Requires specific transformers versions, with a note about a "companion branch" suggesting potential version sensitivity. Deployment and quantization may need specialized tools (e.g., AngelSlim) or complex framework setup (TensorRT-LLM, vLLM, SGLang). The absence of a clear license is a significant adoption blocker.