Abstract
Tencent’s Hunyuan Team presents HY-MT1.5, a family of small-scale machine translation (MT) models with 1.8B and 7B parameters. These models address the "quality-efficiency" deadlock in MT by delivering performance comparable to ultra-large models while maintaining the speed required for real-time applications. Through a novel five-stage training pipeline—incorporating on-policy distillation and rubric-based reinforcement learning—HY-MT1.5-1.8B outperforms models 40x its size, and HY-MT1.5-7B rivals state-of-the-art closed-source systems.
Introduction
Current machine translation faces a significant trade-off: high-accuracy closed-source models (e.g., Gemini-3.0-Pro, GPT-4) are often too slow and costly for high-throughput edge deployment, while lightweight open-source models frequently lack the necessary quality and flexibility. Furthermore, rigid MT systems often struggle with domain-specific constraints like glossary adherence and document structure preservation. HY-MT1.5 bridges this gap, offering "ultra-large" performance in "ultra-small" packages with prompt-driven flexibility.
Methodology
The development of HY-MT1.5 follows a sophisticated five-stage training regime designed to maximize the potential of smaller architectures.
- MT-oriented Pre-training: Continuous pre-training on massive multilingual datasets.
- Supervised Fine-Tuning (SFT): Instruction tuning to ensure the model follows specific translation commands and scenarios.
- Reinforcement Learning (RL): Utilizing Group Relative Policy Optimization (GRPO) with a Rubrics-based Evaluation System. Translations are scored across five dimensions: Accuracy, Fluency, Consistency, Cultural Appropriateness, and Readability.
- Strong-to-Weak On-Policy Distillation: The 7B model acts as a teacher for the 1.8B student. By generating its own data and having the teacher correct the output distribution via reverse KL divergence, the 1.8B model inherits the reasoning capabilities of the larger model.
- Final RL Refinement: A concluding round of RL to polish the distilled 1.8B model.
To ensure deployment readiness, the team implemented quantization techniques, including W8A8C8-FP8 and a 2-bit Quantization-Aware Training (QAT) approach, minimizing performance loss on resource-constrained hardware.
Evaluation and Results
The models were rigorously tested across standard and specialized benchmarks, including Flores-200, WMT25, and a custom Mandarin-Minority test set (covering Tibetan, Mongolian, Uyghur, and Kazakh).
Performance vs. Size
The HY-MT1.5-1.8B model demonstrates exceptional efficiency, outperforming significantly larger models such as Tower-Plus-72B and Qwen3-32B. It achieves approximately 90% of Gemini-3.0-Pro's performance levels. Meanwhile, the HY-MT1.5-7B model reached the 95th percentile of Gemini-3.0-Pro on Flores-200 and surpassed it on the WMT25 and Mandarin-Minority sets.
Speed and Efficiency
Efficiency is a core strength of the HY-MT1.5 family. The 1.8B model maintains an average response time of 0.18 seconds, making it ideal for real-time customer service and instant messaging. FP8 quantization showed negligible performance degradation, confirming the models' suitability for edge deployment.
Prompt-Driven Flexibility
Unlike traditional "black-box" MT systems, HY-MT1.5 supports scenario-based tasks through prompt engineering. This allows for precise terminology intervention (glossary injection) and format preservation without the need for specialized retraining.
Conclusion
HY-MT1.5 proves that sophisticated training methodologies, particularly on-policy distillation and multi-dimensional RL, can enable small models to punch far above their weight class. Future work will focus on expanding language coverage to more low-resource languages and further optimizing the distillation process to push the boundaries of the performance-to-parameter ratio.
Summary updated