tencent has recently officially open-sourced a compact ai translation model called hy-mt1.5-1.8b-1.25bit. the company claims that, while maintaining high performance, the model can run entirely offline on smartphones. currently, hy-mt supports 33 languages—including chinese, english, german, french, japanese, tibetan, and mongolian—as well as five dialects, covering 1,056 translation directions, and has cumulatively won 30 championships in international machine translation competitions.
the core of this technological breakthrough lies in a “radical compression” approach. by employing quantization techniques that use only 1.25 bits per parameter, the model’s size has been slashed from 3.3 gb to 440 mb—about 25% smaller than the previous 1.67-bit solution—while inference speed has increased by roughly 10%, with no loss in quality. in standard benchmark tests, the 440-mb hy-mt delivers translation quality comparable to that of commercial translation services and ultra-large models such as qwen3-32b, effectively bridging the gap between tiny models and those weighing hundreds of gigabytes. this achievement means users no longer need an internet connection or to upload sensitive data to the cloud; they can enjoy high-precision translation right on their phones.
at present, tencent has released an android demo app (in apk format) that supports cross-screen offline translation of text within any app on the device. users can directly invoke the model in social, reading, or office scenarios without having to switch apps. industry observers note that, with google’s launch of the on-device model gemma4, on-device ai has become the new frontier of technological competition. tencent’s hy-mt, through its breakthrough in quantization technology, has dramatically lowered the computational requirements for high-quality ai translation, providing a highly competitive open-source foundation for on-device ai applications in complex privacy-sensitive environments. industry analysts believe that this “small but powerful” model strategy will accelerate the migration of ai capabilities from the cloud to individual devices, further unlocking the potential of mobile intelligence.
