Tencent HY-MT1.5: Advanced Translation Models Launched

Overview

Tencent Hunyuan researchers have released HY-MT1.5, a multilingual machine translation family targeting both mobile devices and cloud systems with identical training recipes and metrics. This model family includes two variants, HY-MT1.5-1.8B and HY-MT1.5-7B, supporting mutual translation across 33 languages with 5 ethnic and dialect variations. Both models are available on GitHub and Hugging Face under open weights.

Model Family and Deployment Targets

HY-MT1.5-7B is an enhancement of the WMT25 championship system Hunyuan-MT-7B, optimized for explanatory translation and mixed language scenarios. It supports terminology intervention, contextual translation, and formatted translation.

Conversely, HY-MT1.5-1.8B is a compact variant. While having less than one-third the parameters of HY-MT1.5-7B, it maintains comparable translation performance across benchmarks. After quantization, the 1.8B model runs on edge devices and enables real-time translations.

The quantized HY-MT1.5-1.8B operates on devices with about 1 GB of memory and achieves an average response time of 0.18 seconds for Chinese inputs of approximately 50 tokens, outperforming mainstream translation APIs. In contrast, HY-MT1.5-7B expects latency around 0.45 seconds for high-quality outputs.

Holistic Training Framework

HY-MT1.5 is defined as a translation-specific language model employing a multi-stage training pipeline:

1. General Pre-training: Initiates with large-scale multilingual text to form shared representations.
2. MT-Oriented Pre-training: Exposes the model to parallel corpora, aligning distribution with translation tasks.
3. Supervised Fine Tuning: Utilizes high-quality parallel data to enhance model accuracy and domain coverage.
4. On Policy Distillation from 7B to 1.8B: HY-MT1.5-7B guides HY-MT1.5-1.8B, yielding a student model that retains key translation behaviors at a lower cost.
5. Reinforcement Learning with Rubric-Based Evaluation: Optimizes both models based on detailed human evaluations over multiple quality metrics.

This unique pipeline distinguishes machine translation development from chat-oriented LLM training.

Benchmark Results Against Open and Commercial Systems

HY-MT1.5 has been evaluated on Flores 200, WMT25, and a Mandarin-to-minority language benchmark:

• Flores 200: HY-MT1.5-7B achieved XCOMET-XXL scores of 0.8690 for ZH to XX, and 0.9093 for EN to XX, outperforming specialized models and competing with larger models like Qwen3-235B-A22B.
• WMT25: Scores reached 0.6159, surpassing other translation models.
• Mandarin to Minority Languages: Scores of 0.6174 were among the highest recorded.

Human evaluation (0-4 scale) for translations saw HY-MT1.5-1.8B achieving an average score of 2.74, superior to other leading systems.

Practical Features for Product Use

HY-MT1.5 provides three key capabilities in production environments:

1. Terminology Intervention: Injects term mappings for controlled translations critical in specialized fields.
2. Context-Aware Translation: Adjusts translations based on contextual cues.
3. Format-Preserving Translation: Maintains structural integrity of text during translation, essential for HTML or XML content.

These features function through prompt formats, available when utilizing public weights.

Quantization and Edge Deployment

HY-MT1.5-1.8B has undergone FP8 and Int4 quantization, enabling efficient deployment:

• FP8 remains close in quality to full precision, while Int4 exhibits some quality drops. Both variants are available for implementation on Hugging Face.

Key Takeaways

• HY-MT1.5 introduces two models supporting translation across 33 languages with open weights shared.
• HY-MT1.5-1.8B is optimal for edge deployment, delivering low latency and high performance.
• HY-MT1.5-7B ranks competitively against larger models while employing a robust training pipeline for efficiency.
• Enhanced features cater to production needs, supporting various deployment environments.