Hy-MT2

› The “Fast-Thinking” multilingual translation model series released by Tencent Hunyuan supports mutual translation across 33 languages and comes in three variants: 1.8B, 7B, and 30B-A3B.

· Developer: Tencent Hunyuan
· Release date: 2026-05-21
· HuggingFace: tencent/Hy-MT2 collection
· GitHub: Tencent-Hunyuan/Hy-MT2
· Research paper: HY-MT1.5 Technical Report (arxiv:2512.24092)


Overview of the Series

Hy-MT2 is a dedicated translation model series developed by Tencent Hunyuan for complex real-world scenarios. Its core feature is the “fast-thinking” mechanism—it generates an internal chain of thought during inference before producing the final translation output. This makes it highly adept at handling intricate translation instructions involving specialized terminology, stylistic nuances, and structured data.

Alongside this series, IFMTBench was also open-sourced; it serves as a benchmark specifically designed to evaluate models’ ability to follow translation instructions.


Specification Comparison

Specification Parameter count Architecture type Active parameters Ideal use cases
Hy-MT2-1.8B 1.8B Dense 1.8B Edge device deployment, low-resource environments
Hy-MT2-7B 7B Dense 7B Balancing performance and resource consumption
Hy-MT2-30B-A3B 30B MoE 3B (per inference) Delivering top-tier translation quality

Performance Highlights

1.8B variant: Outperforms mainstream commercial APIs such as Microsoft Translator and Doubao across overall metrics.

7B / 30B-A3B variants: Under fast-thinking mode, they surpass open-source models like DeepSeek-V4-Pro and Kimi K2.6.

• Comprehensive evaluation across four key dimensions: general translation, real-world business scenarios, domain-specific expertise, and instruction adherence.


How to Choose the Right Specification

Do you need edge/mobile deployment?
  └─ Yes → Go for Hy-MT2-1.8B (the 1.25-bit quantized version requires just 440MB of storage).

Is your GPU VRAM less than 20GB?
  └─ Yes → Hy-MT2-7B is the ideal choice.

Are you aiming for peak translation quality and have ample resources available?
  └─ Yes → Opt for Hy-MT2-30B-A3B (an MoE model where only 3B parameters activate per inference).

Supported Languages (38 total)

Chinese, English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, Malay, Indonesian, Filipino, Hindi, Traditional Chinese, Polish, Czech, Dutch, Khmer, Burmese, Persian, Gujarati, Urdu, Telugu, Marathi, Hebrew, Bengali, Tamil, Ukrainian, Tibetan, Kazakh, Mongolian, Uyghur, Cantonese.


All Available Models

Model name Description HuggingFace link
Hy-MT2-1.8B Base 1.8B model Link
Hy-MT2-1.8B-FP8 1.8B model with FP8 quantization Link
Hy-MT2-1.8B-GGUF 1.8B model in GGUF format (compatible with llama.cpp) Link
Hy-MT2-1.8B-2bit-GGUF 1.8B model with 2-bit GGUF quantization Link
Hy-MT2-1.8B-1.25bit-GGUF 1.8B model with 1.25-bit GGUF quantization (extreme edge optimization) Link
Hy-MT2-7B Base 7B model Link
Hy-MT2-7B-FP8 7B model with FP8 quantization Link
Hy-MT2-7B-GGUF 7B model in GGUF format (compatible with llama.cpp) Link
Hy-MT2-30B-A3B Base 30B MoE model Link
Hy-MT2-30B-A3B-FP8 30B MoE model with FP8 quantization Link

Version History

Version Release date Key updates
Hy-MT2 2026-05-21 Introduced 1.8B, 7B, and 30B-A3B variants with fast-thinking mechanism; IFMTBench was also open-sourced simultaneously.
Hy-MT1.5 2025-12-30 Released 1.8B and 7B variants.
Hunyuan-MT 2025-09-01 Launched the first-generation series featuring 7B and Chimera-7B models.

Related Resources

Explanation of the IFMTBench benchmark

Model training guide

AngelSlim quantization tool

• Technical support: [email protected]