Hy-MT2-7B

1. Multilingual Translation

Hy-MT2 supports translation across 33 languages.

That includes major languages such as Chinese, English, Japanese, Korean, French, German, and Spanish, along with languages such as Uyghur and Tibetan.

The model was trained specifically around translation rather than treating translation as one feature inside a general-purpose chatbot.

2. Instruction Following

This is one of the more useful parts of Hy-MT2.

You can give the model a specific rule such as:

“Keep the translation concise and limit each sentence to 15 words.”

It will try to follow that instruction while translating.

That matters in production work. Translation jobs often come with extra constraints: certain terms must stay fixed, punctuation cannot change, formatting has to remain intact, or some fields must not be translated at all.

General-purpose models can start dropping those requirements once the task gets complicated.

Hy-MT2 was trained with this kind of control in mind. Tencent also released IFMTBench, a benchmark designed to measure how well translation models follow complex instructions.

3. Domain-Specific Translation

Hy-MT2 has been optimized for areas such as finance, politics, and education.

On DomainMTBench, the 7B model improved its GEMBA score from 92.04 on Hy-MT1.5 to 92.79.

For professional translation, fluency is only part of the job. Terminology needs to stay consistent across an entire document, especially in financial reports, technical material, and industry-specific content.

Hy-MT2 is designed with that kind of consistency in mind.

4. Heavy Quantization and On-Device Use

Hy-MT2 supports FP16, 8-bit, 4-bit, 2-bit, and even 1.25-bit quantization.

The 1.8B model can be compressed to roughly 440MB and run locally on mobile hardware.

That opens up some practical use cases: offline translation, privacy-sensitive applications, and mobile apps that cannot rely on a cloud connection.

Translation is also a good fit for smaller models because the task itself is relatively well defined.

5. What Changed From Hy-MT1.5

On FLORES-200, Hy-MT2-7B increased its XCOMET-XXL score from 80.98 to 86.89.

It also outperformed Hy-MT1.5 across all three reported metrics on WMT25.

The improvements show up in everyday use too. Translations read more naturally, long passages stay more consistent toward the end, formatting and terminology instructions are missed less often, and complex prompts are handled more reliably.

None of those changes are flashy on their own. Together, they make the model much easier to use in a real translation workflow.

Strengths

Hy-MT2-7B is competitive for its size, with a FLORES-200 score of 86.89, reported as roughly 97.9% of Gemini 3.1 Pro's score on the same comparison.

It offers fine-grained control over terminology, formatting, and output constraints.

You can either self-host the open-source model or use Tencent Cloud's managed API.

The quantized versions also make it practical to run on consumer GPUs and, with the smaller 1.8B model, on mobile hardware.

Limitations

It is a translation model, not a general AI assistant.

If you need coding, web search, general writing, or open-ended conversation, Hy-MT2-7B is not the right tool.

Self-hosting still requires some technical knowledge. You need to manage the GPU environment, inference framework, and quantization setup yourself.

The ecosystem and tooling outside Chinese-language use cases are also less mature than what you get with larger general-purpose models.

Recommended for

Developers and teams working on localization, cross-border products, professional document translation, or API replacement projects.

It is also a good fit for applications that need local deployment or offline translation.

Not recommended for

People who only translate a few short sentences from time to time and do not care much about terminology, formatting, or consistency.

If what you really want is one model that can chat, write, code, search, and translate, this is not it.