Tencent Releases HY-MT1.5 Translation Models: 1.8B & 7B Parameters for Cloud & Edge

Tencent HY-MT1.5: A New Translation Models

Tencent researchers have released HY-MT1.5, a new family of multilingual machine translation models comprising 1.8B and 7B parameter versions. These models are designed for both on-device and cloud deployment, supporting 33 languages and 5 variations, and are available with open weights on GitHub and Hugging Face.

Why This Matters

Current large language models (LLMs) often struggle with efficient translation, demanding significant computational resources and latency. HY-MT1.5 addresses this by offering a distilled 1.8B parameter model capable of running on edge devices with limited memory, while still achieving competitive translation quality – a critical step towards widespread, low-cost machine translation access.

Key Insights

• Flores 200 scores (HY-MT1.5-7B): Reached 0.8690 for ZH to XX, surpassing iFLYTEK Translator, 2026.
• Distillation for efficiency: HY-MT1.5-1.8B is trained using a teacher-student approach, inheriting capabilities from the larger 7B model with reduced computational cost.
• Prompt-driven features: Terminology intervention, context-aware translation, and format preservation are enabled through specific prompt templates, simplifying integration into existing systems.

Working Example

# Example prompt for terminology intervention (Python)
prompt = """Translate the following sentence, ensuring "混元珠" is translated as "Chaos Pearl":
Sentence: The explorer discovered a 混元珠 in the ancient temple.
"""
# This prompt would be sent to the HY-MT1.5 model.

Practical Applications

• Mobile Translation Apps: HY-MT1.5-1.8B can power real-time translation directly on smartphones with minimal latency.
• Pitfall: Relying solely on LLM size; HY-MT1.5 demonstrates that specialized training and distillation techniques can outperform larger, general-purpose models in specific tasks like translation.

References:

• https://www.marktechpost.com/2026/01/04/tencent-researchers-release-tencent-hy-mt1-5-a-new-translation-models-featuring-1-8b-and-7b-models-designed-for-seamless-on-device-and-cloud-deployment/
• [GitHub Repo] (link not explicitly provided in context)
• [Model Weights on HF] (link not explicitly provided in context)