The breakthrough lies in aggressive 1.25-bit quantization, shrinking the model from 3.3GB to 440MB, about 25% smaller than the prior 1.67-bit scheme, with roughly 10% faster inference and no quality loss.

Tencent open-sources Hy-MT1.5-1.8B-1.25bit, a compact AI translation model designed to run completely offline on smartphones.

Hy-MT supports 33 languages and five dialects, covering 1056 translation directions, and has claimed strong performance with 30 international machine translation championships.

Tencent released an Android APK demonstration app enabling cross-screen offline translation within any mobile application.

Benchmark tests show the 440MB Hy-MT delivers translation quality on par with commercial services and large models like Qwen3-32B despite its small size.

Industry observers say on-device AI is a competitive frontier, and Hy-MT’s quantization lowers the computing power bar for high-quality offline translation in privacy-conscious settings, alongside rivals like Google Gemma4.