Hy-MT2-1.8B-1.25Bit-GGUF

<p align="left"> English&nbsp;｜&nbsp;<a href="https://huggingface.co/AngelSlim/Hy-MT2-1.8B-1.25Bit-GGUF/blob/main/README_CN.md">中文</a> </p> <br> <p align="center"> <img src="logo-en.png" width="400"/> <br> </p> <div align="center" style="line-height: 1;">

  

</div> <p align="center"> 🖥️&nbsp;<a href="https://aistudio.tencent.com/llm/en?tabIndex=0"><b>Official Website</b></a>&nbsp;&nbsp;|&nbsp;&nbsp; 💬&nbsp;<a href="https://github.com/Tencent-Hunyuan/Hy-MT2"><b>GitHub</b></a>&nbsp;&nbsp;|&nbsp;&nbsp; 🪡&nbsp;<a href="https://github.com/Tencent/AngelSlim/tree/main"><b>AngelSlim</b></a></p>

Model Introduction

Hy-MT2-1.8B-1.25 Bit-GGUF produced by AngelSlim, For more detailed information, please refer to [AngelSlim].

Hy-MT2 is a family of “fast-thinking” multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. For on-device deployment, AngelSlim 1.25-bit extreme quantization reduces the storage requirement of the 1.8B model to only 440 MB and improves inference speed by 1.5x. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B-A3B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.

In this release, we also open-source IFMTBench, a benchmark for evaluating translation instruction-following capabilities.

We also welcome everyone to use our released Hy-MT2-Translator Skill, which makes it easy to integrate Hy-MT2 series models for translation tasks. Download links: ClawHub and SkillHub.

Now, Tencent Hy is officially partnering with WMT26 for the "Video Subtitle Translation Task" (https://www2.statmt.org/wmt26/video-subtitle-translation.html). Participants who use the Hy-MT model series to compete in the "General Machine Translation Task" (https://www2.statmt.org/wmt26/translation-task.html) and the "Video Subtitle Translation Task" will have the chance to win special awards sponsored by Hunyuan. We sincerely invite everyone to participate and jointly push the boundaries of machine translation technology!

News

• 2026.5.21 We open-sourced Hy-MT2-1.8B/Hy-MT2-7B/Hy-MT2-30B-A3B/IFMTBench on HuggingFace and ModelScope.
• 2025.12.30 We open-sourced HY-MT1.5-1.8B and HY-MT1.5-7B on HuggingFace and ModelScope.
• 2025.9.1 We open-sourced Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B on HuggingFace and ModelScope.

Results

<div align='center'> <img src="main_result.png" width = "100%" /> </div>

For more experimental results and analysis, please refer to our report.

 

Model Links

Hy-MT2 Translation Task Instruction Examples (Chinese-English Comparison)

Note: In the following examples, both source_lang and target_lang should use the full language names. Chinese names should be used in Chinese prompts, and English names should be used in English prompts.

---

Inference and Deployment

transformers

transformers>=5.6.0

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "tencent/Hy-MT2-30B-A3B"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Load model
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

model.eval()

# Example inference
prompt = "将以下文本翻译成英语,注意只需要输出翻译后的结果,不要额外解释:\n\n今天天气真好。"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=4096,
    )
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)

vllm

Build vLLM from source:

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate
git clone https://github.com/vllm-project/vllm.git
cd vllm
uv pip install --editable . --torch-backend=auto

Start the vLLM server:

vllm serve tencent/Hy-MT2-30B-A3B --tensor-parallel-size 1

sglang

Build SGLang from source:

git clone https://github.com/sgl-project/sglang
cd sglang
pip3 install pip --upgrade
pip3 install "transformers>=5.6.0"
pip3 install -e "python"

Launch SGLang server:

python3 -m sglang.launch_server --model tencent/Hy-MT2-30B-A3B --tp 1

llama_cpp

❕❕ This gguf depends on our STQ kernel, which is released at PR #22836.

Clone llama.cpp

git clone https://github.com/ggml-org/llama.cpp.git

Enter the llama.cpp folder

cd llama.cpp

Build llama.cpp

cmake -B build
cmake --build build --config Release

Run a completion example

./build/bin/llama-completion \
  --model model.gguf  \
  -p "Translate the following segment into Chinese, without additional explanation：Hello" \
  --jinja \
  -ngl 0 \
  -n 64 -st 

Run the llama.cpp benchmark

./build/bin/llama-bench -m model_zoo/model.gguf  -ngl 0

For 1.8B and 7B, we recommend using the following parameters for inference. Note that our models do not have a default system_prompt.

{
  "temperature": 0.7,
  "top_p": 0.6,
  "top_k": 20,
  "repetition_penalty": 1.05,
  "max_tokens": 4096
}

For 30B-A3B, we recommend using the following parameters for inference. Note that our models do not have a default system_prompt.

{
  "temperature": 0.7,
  "top_p": 1.0,
  "top_k": -1,
  "repetition_penalty": 1.0,
  "max_tokens": 4096
}

Model Training

Hy-MT2 provides a complete model training pipeline, supporting both full-parameter fine-tuning and LoRA fine-tuning, as well as multiple DeepSpeed ZeRO configurations and LLaMA-Factory integration.

For detailed training documentation, please refer to: Model Training Guide

Quantization Tool

We provide AngelSlim, an easy-to-use, comprehensive, and efficient large model compression toolkit covering common quantization algorithms, low-bit quantization, speculative sampling, and more.

Supported Languages

Citing Hy-MT2

@misc{hy-mt1.5,
      title={HY-MT1.5 Technical Report}, 
      author={Mao Zheng and Zheng Li and Tao Chen and Mingyang Song and Di Wang},
      year={2025},
      eprint={2512.24092},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.24092}, 
}

Contact Us

If you would like to leave feedback for our R&D and product teams, you are welcome to contact the Tencent Hunyuan LLM team. You can reach us by email at [email protected].