Tencent Hunyuan HY-MT 1.5: 33-Language Translation Model, Small Size Big Results

What Is This?

Tencent Hunyuan recently open-sourced HY-MT 1.5, a large language model specifically designed for translation. It supports 33 languages, including Cantonese and Traditional Chinese dialects.

Most impressively, this model is an upgraded version of the WMT25 (top-tier machine translation competition) champion model, delivering exceptional translation quality.

Project URL: https://github.com/Tencent-Hunyuan/HY-MT

Key Highlights

Small Model, Big Performance

HY-MT offers two versions:

Here’s the key point: The 1.8B small model performs almost as well as the 7B version. With only one-third of the parameters, translation quality is nearly identical - great news for developers looking to run translation on mobile or edge devices.

Three Practical New Features

1. Terminology Intervention - Specify how professional terms should be translated (medical, legal terminology, etc.)
2. Context-Aware Translation - Uses context for more coherent translations
3. Format-Preserving Translation - Maintains original document formatting, convenient for document translation

Language Coverage

Supports 33 languages including Chinese, English, Japanese, Korean, French, German, Russian, Arabic, and more. Plus 5 additional dialect/variant options (Cantonese, Traditional Chinese, etc.).

Model Downloads

Tencent provides multiple quantized versions for those with limited GPU memory:

HY-MT1.5-1.8B          # Base version
HY-MT1.5-1.8B-FP8      # FP8 quantization
HY-MT1.5-1.8B-GPTQ-Int4 # Int4 quantization, lowest memory usage

HY-MT1.5-7B            # Standard version
HY-MT1.5-7B-FP8        # FP8 quantization
HY-MT1.5-7B-GPTQ-Int4  # Int4 quantization

How to Use?

Transformers Inference

The simplest approach - just install transformers and run:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "tencent/HY-MT1.5-7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Translate the following Chinese to English: Artificial intelligence is changing our way of life."

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    top_k=20,
    top_p=0.6,
    repetition_penalty=1.05,
    temperature=0.7,
    max_new_tokens=256
)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Official recommended parameters: top_k=20, top_p=0.6, repetition_penalty=1.05, temperature=0.7

Docker One-Click Deployment

For those who don’t want to deal with environment setup, use the official Docker image:

docker pull hunyuaninfer/hunyuan-7b:hunyuan-7b-trtllm

docker run -d --gpus all \
  -p 8000:8000 \
  hunyuaninfer/hunyuan-7b:hunyuan-7b-trtllm

Supported Inference Frameworks

Besides Transformers, these mainstream frameworks are also supported:

• TensorRT-LLM - For maximum performance
• vLLM - High-concurrency service deployment
• SGLang - Flexible inference engine

Want to Fine-tune?

Use the LLaMA-Factory framework:

# Single machine training
llamafactory-cli train examples/hy-mt/sft.yaml

# Multi-node distributed training
torchrun --nproc_per_node=8 \
  llamafactory-cli train examples/hy-mt/sft.yaml

For example, if you want to optimize translation for medical documents, just prepare your dataset and fine-tune.

Ideal Use Cases

• Real-time Translation Apps - 1.8B version is sufficient and fast
• Document Translation - Format-preserving translation maintains layout
• Professional Domains - Terminology intervention ensures accuracy
• Translation Services - 7B + vLLM for high-throughput deployment
• Multilingual Customer Service - 33 languages for instant translation

Where to Download?

• Hugging Face: tencent/HY-MT1.5-7B
• ModelScope: Search for HY-MT on ModelScope
• GitHub: Tencent-Hunyuan/HY-MT

Summary

As an upgraded version of the WMT champion model, HY-MT 1.5 has several clear advantages:

• 1.8B small model approaches 7B performance, deployment-friendly
• Terminology intervention and context-aware translation are practical features
• 33 languages + dialects for broad coverage
• Multiple quantized versions and inference framework support

If you’re working on multilingual projects, this model is worth trying.