Tencent HY-MT1.5: Bridging Multilingual Translation Gaps with Efficient Edge and Cloud Models

In a world where seamless communication across languages is essential for global collaboration, how can AI translation models balance high performance with accessibility on resource-constrained devices?

Advancements in Multilingual Machine Translation

Tencent’s Hunyuan research team has introduced HY-MT1.5, a family of multilingual machine translation models designed to support both on-device and cloud-based deployments. This release features two variants: the HY-MT1.5-7B, a larger model optimized for high-quality server applications, and the HY-MT1.5-1.8B, a compact version tailored for edge devices. Both models facilitate mutual translation across 33 languages, including five ethnic and dialect variations, addressing a broad spectrum of linguistic needs in diverse regions. The HY-MT1.5-7B builds on the foundation of the previous WMT25 championship system, Hunyuan-MT-7B, with enhancements for explanatory translations, mixed-language processing, and specialized features like terminology intervention and contextual formatting.

These improvements enable more nuanced handling of complex scenarios, such as domain-specific terminology in legal or medical texts. In contrast, the HY-MT1.5-1.8B, with fewer than one-third the parameters of its counterpart, achieves comparable benchmark performance after quantization, making it viable for real-time applications on mobile hardware. This dual-model approach reflects a growing trend in AI development toward scalable architectures that reduce dependency on cloud infrastructure, potentially lowering latency and costs for end-users in bandwidth-limited areas. By releasing open weights on platforms like GitHub and Hugging Face, Tencent democratizes access, allowing developers worldwide to integrate these models into applications without proprietary barriers.

Training Pipeline and Architectural Innovations

The development of HY-MT1.5 employs a multi-stage training framework tailored specifically to machine translation tasks, diverging from general-purpose large language model pipelines. This holistic process ensures alignment with translation objectives, emphasizing accuracy, fluency, and cultural appropriateness. Key stages include:

• General pre-training: Exposure to large-scale multilingual text corpora to establish shared linguistic representations across languages.
• MT-oriented pre-training: Alignment with parallel corpora and translation-specific objectives to shift focus from open-ended generation to precise bilingual tasks.
• Supervised fine-tuning: Refinement using high-quality sentence- and document-level parallel data, enhancing literal accuracy, domain coverage, and directionality (e.g., Chinese-to-English versus English-to-Chinese).
• On-policy distillation: Transfer of knowledge from the 7B teacher model to the 1.8B student via approximately 1 million monolingual prompts, utilizing reverse Kullback-Leibler divergence to preserve translation behaviors at reduced computational cost.
• Reinforcement learning with rubric-based evaluation: Optimization using a group relative policy algorithm, guided by a reward model derived from human assessments on axes like accuracy, fluency, idiomaticity, and cultural fit.

Benchmark Performance and Deployment Implications

Evaluations of HY-MT1.5 on standard benchmarks like Flores 200, WMT25, and a Mandarin-to-minority language dataset demonstrate competitive results against both open-source and commercial systems. Metrics such as XCOMET-XXL and CometKiwi were used to assess translation quality. Notable statistics include:

• On Flores 200, the 7B model scores 0.8690 for Chinese-to-other languages (ZH-XX), 0.9093 for English-to-other (EN-XX), and 0.8098 for other-to-other (XX-XX), outperforming specialized translators like iFLYTEK and Doubao while matching mid-sized general models such as Qwen3-235B-A22B.
• For WMT25, the 7B achieves 0.6159 on XCOMET-XXL, surpassing Gemini 3.0 Pro by 0.065 and exceeding models like Seed-X-PPO-7B and Tower-Plus-72B; the 1.8B variant scores 0.5308, still above many peers.
• In Mandarin-to-minority translations, the 7B reaches 0.6174, topping baselines including Gemini 3.0 Pro, with the 1.8B at 0.5806, outperforming larger models like DeepSeek-V3.2.