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

Why It Matters

By delivering high‑quality, low‑latency translation on consumer hardware, HY‑MT1.5 lowers barriers for multilingual applications and intensifies competition with major cloud translation services.

Key Takeaways

• HY-MT1.5 supports 33 languages plus 5 dialects
• 1.8B model runs on 1 GB memory, 0.18 s latency
• 7B model matches WMT25 champion, near‑state‑of‑the‑art
• Prompt templates enable terminology, context, format preservation
• Open weights on GitHub and Hugging Face accelerate adoption

Pulse Analysis

The translation market has long been dominated by cloud‑only services that require constant connectivity and high‑cost infrastructure. As mobile AI chips become more capable, enterprises are seeking on‑device solutions that can translate instantly without sacrificing privacy. HY‑MT1.5 addresses this shift by offering a lightweight 1.8 B model that fits within a gigabyte of RAM, delivering sub‑200 ms responses for typical Chinese inputs. This performance level narrows the gap between edge AI and traditional server‑based translators, opening new use cases in travel, education, and real‑time communication.

Tencent’s edge comes from a holistic training pipeline that blends large‑scale multilingual pre‑training, translation‑specific objectives, supervised fine‑tuning, and a teacher‑student distillation stage. The 7 B model serves as a teacher, passing nuanced translation behavior to the smaller student via on‑policy distillation and reverse KL divergence. A final reinforcement‑learning phase, guided by human‑rated rubrics, refines accuracy, fluency, and cultural appropriateness. Post‑training quantization to FP8 and Int4 further compresses the model, preserving most of its XCOMET‑XXL scores while enabling deployment on consumer devices.

Open‑source availability amplifies HY‑MT1.5’s impact. By publishing weights on Hugging Face and providing GGUF formats, Tencent invites developers to integrate the models into existing LLM stacks, experiment with custom prompts, and extend language coverage. The built‑in capabilities for terminology injection, context‑aware translation, and format preservation make the models production‑ready for domains such as legal, medical, and e‑commerce. As the models demonstrate competitive scores against Google, Microsoft, and emerging Gemini systems, they signal a democratization of high‑quality machine translation and set a new benchmark for on‑device multilingual AI.