On standard benchmarks, it matches the performance of commercial translation services and significantly larger models such as Qwen3-32B. This is made possible through compression to 1.25 bits per parameter: memory requirements drop from 3.3 GB to just 440 MB — approximately 25% smaller and 10% faster than earlier 1.67-bit approaches, with no loss in quality.

At just 440 MB, Hy-MT1.5-1.8B-1.25bit delivers translation quality on par with models that are several hundred gigabytes in size.
At just 440 MB, Hy-MT1.5-1.8B-1.25bit delivers translation quality on par with models that are several hundred gigabytes in size.

Tencent also offers an Android demo app available as an APK download, which translates words in the background across all apps without requiring an internet connection. According to the company, the model has won 30 first-place rankings at international machine translation competitions. Google is pursuing a similar direction with Gemma 4, which also runs local AI directly on mobile devices.