Hunyuan Translation Model Version 1.5 (HY-MT1.5) is a suite of machine translation models open-sourced by Tencent, featuring two primary variants: HY-MT1.5-1.8B and HY-MT1.5-7B. Both models are designed for mutual translation across 33 languages, including common global languages and 5 ethnic/dialect variations, supporting a total of 38 language pairs.

HY-MT1.5-1.8B: This is a 1.8 billion parameter translation model. It boasts industry-leading performance for its size, often surpassing most commercial translation APIs. A key advantage is its ability to achieve high speed and high quality simultaneously. After quantization (available in FP8 and GPTQ-Int4 versions), the 1.8B model can be deployed efficiently on edge devices, making it suitable for real-time translation scenarios. The base model utilizes BF16 precision for its parameters.

HY-MT1.5-7B: This is a larger 7 billion parameter model, representing an upgraded version of Tencent's WMT25 championship model. While having significantly more parameters than the 1.8B version, the 1.8B model still delivers comparable translation performance. The 7B model is specifically optimized for complex translation scenarios, including explanatory translation and mixed-language content. It also offers advanced features such as terminology intervention, contextual translation, and formatted translation. Quantized versions (FP8 and GPTQ-Int4) are also available for the 7B model.

Core Methodology and Features:
The models operate as large language models (LLMs) fine-tuned for translation tasks. Their multi-lingual capability is achieved through extensive training on a diverse dataset covering the 33 supported languages and their various combinations. The specific advancements in HY-MT1.5-7B, such as optimization for explanatory and mixed-language scenarios, suggest sophisticated prompt engineering and possibly specialized training data or architectural adjustments to handle such complex linguistic structures.

A core aspect of interaction with these models is the use of specific prompt templates to guide the translation process and leverage their advanced features:

• Standard Translation (ZH<=>XX): 将以下文本翻译为{target_language}，注意只需要输出翻译后的结果，不要额外解释： {source_text}
• Standard Translation (XX<=>XX, excluding ZH<=>XX): Translate the following segment into {target_language}, without additional explanation. {source_text}
• Terminology Intervention: This feature allows users to provide specific translations for key terms. The template incorporates both the desired term translation and the main text: 参考下面的翻译： {source_term} 翻译成 {target_term} 将以下文本翻译为{target_language}，注意只需要输出翻译后的结果，不要额外解释： {source_text}
• Contextual Translation: This enables the model to consider preceding text for better translation of the current segment, without translating the context itself: {context} 参考上面的信息，把下面的文本翻译成{target_language}，注意不需要翻译上文，也不要额外解释： {source_text}
• Formatted Translation: Designed to preserve formatting tags (e.g., $<sn></sn>$) from the source text into the translated output. The model is instructed to maintain these tags and output in a specific XML-like format: $将以下<source></source>之间的文本翻译为中文，注意只需要输出翻译后的结果，不要额外解释，原文中的<sn></sn>标签表示标签内文本包含格式信息，需要在译文中相应的位置尽量保留该标签。输出格式为：<target>str</target> <source>{src_text_with_format}</source>$

Inference Parameters:
The recommended inference parameters for generating outputs are:

• top_k: 20
• top_p: 0.6
• repetition_penalty: 1.05
• temperature: 0.7

The model does not utilize a default system prompt.

Deployment and Compatibility:
The models are compatible with the Hugging Face transformers library (v4.56.0 recommended). For loading FP8 quantized models, a specific modification is required: changing "ignored_layers" to "ignore" in config.json and upgrading compressed-tensors to version 0.11.0. This highlights a dependency on specific library versions and configuration for optimal performance with quantized models.

The open-sourcing timeline indicates a continuous development effort, with the 1.5 versions released on December 30, 2025, following earlier releases of Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B on September 1, 2025. The technical report (arXiv:2512.24092) provides further experimental results and analysis.