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

2026-01-05 · Source: MarkTechPost · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Intermediate, medium

Summary

Tencent Hunyuan researchers have released HY-MT1.5, a new multilingual machine translation family comprising two models: HY-MT1.5-1.8B and HY-MT1.5-7B. These models are designed for both mobile devices and cloud systems, supporting mutual translation across 33 languages with 5 ethnic and dialect variations. Available on GitHub and Hugging Face with open weights, HY-MT1.5-7B is an upgraded version of the WMT25 championship system, optimized for explanatory and mixed-language scenarios, and includes native support for terminology intervention, contextual translation, and formatted translation. The compact HY-MT1.5-1.8B, with less than one-third the parameters, delivers comparable performance and, after quantization, can run on edge devices with approximately 1 GB of memory, achieving an average response time of 0.18 seconds for 50-token Chinese inputs while surpassing mainstream commercial APIs in quality. Both models are trained using a holistic, multi-stage pipeline that includes general pre-training, MT-oriented pre-training, supervised fine-tuning, on-policy distillation from 7B to 1.8B, and reinforcement learning with rubrics-based evaluation.

Key takeaway

For AI Engineers and Machine Learning Engineers deploying machine translation, HY-MT1.5 offers a compelling open-weight solution. Your teams can leverage the HY-MT1.5-1.8B model for efficient edge deployment on devices with 1 GB memory, achieving high quality and low latency. Consider integrating its prompt-driven features for terminology control, context-aware translation, and format preservation to enhance production system robustness and accuracy.

Key insights

Tencent's HY-MT1.5 offers high-quality, open-weight machine translation models optimized for diverse deployment scenarios.

Principles

• Distillation enables smaller models to inherit larger model behavior.
• Rubrics-based RL fine-tuning improves translation quality.
• Prompt-driven features enhance production system utility.

Method

A multi-stage pipeline combines general and MT-oriented pre-training, supervised fine-tuning, on-policy distillation, and reinforcement learning with rubrics-based human evaluation to optimize translation models.

In practice

• Use HY-MT1.5-1.8B for on-device, low-latency translation.
• Employ prompt templates for terminology, context, and format control.
• Utilize FP8 or Int4 quantization for memory-constrained deployment.

Topics

• Multilingual Machine Translation
• On-Device AI
• Model Distillation
• Reinforcement Learning
• Prompt Engineering

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, NLP Engineer

Related on AIssential

• ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation — ReflectMT internalizes reflection to achieve high-quality, efficient machine translation without explicit reasoning steps during inference.
• Surpassing Scale by Efficiency: A Compact 135M Parameter Foundational LLM Natively Adapted for the Bangla Language — A compact 135M parameter LLM for Bangla achieves 1B-tier performance by efficiently merging tokens, overcoming low-resource script challenges.
• Celebrating 20 years of Google Translate: Fun facts, tips and new features to try — Google Translate's 20-year evolution highlights AI's role in breaking language barriers and fostering global connection.
• Cohere Launches Tiny Multilingual Open Weight Model — Tiny Aya offers open-weight, multilingual AI models to enhance linguistic diversity and edge device accessibility.
• Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild — Hy-MT2 delivers high-quality, efficient multilingual translation with strong instruction-following across diverse real-world and resource-constrained scenarios.
• How Roblox Uses AI to Translate 16 Languages in 100 Milliseconds — A single Mixture of Experts model can efficiently handle real-time translation across many language pairs at massive scale.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MarkTechPost.