Tencent has released an offline translation model for mobile, have a look?

Tencent has announced 'Hy-MT1.5-1.8B-1.25bit', an offline AI translation model that runs on smartphones without an internet connection. The model was released on April 29 through 'AngelSlim', Tencent's AI model compression project, and is currently available on GitHub and Hugging Face along with an Android demo app.

Most existing translation services work by communicating with a server, but this model handles translation entirely on the smartphone itself. That means once the model is downloaded, you can use translation without an internet connection, and there's no need to send the sentences you want translated to an external server.

It supports 33 languages, including Korean, English, Chinese, and Japanese, with 1,056 translation directions. It also uses inference optimized for mobile CPUs, so translation works without a separate cloud server.

Tencent has shared a demo of the model running on a smartphone with a Snapdragon 865 processor and 8GB of RAM. The Android demo app also includes a feature that extracts text from other apps while you're using them and translates it.

Follow-up model Hy-MT2 also released

Hy-MT2 is also an ultra-lightweight model of about 440MB, and it has been confirmed to improve translation accuracy and the handling of complex translation instructions.

The Android demo app currently released runs on the existing Hy-MT1.5. To run the Hy-MT2 model in a mobile environment, a separate compatible inference environment must be set up first.
Also, under the license policy, the community license for the existing Hy-MT1.5 model excludes South Korea from the permitted regions, so use within Korea requires special caution.

On the other hand, the follow-up model Hy-MT2 1.8B has been released under the Apache 2.0 license.
As on-device AI spreads, being able to do multilingual translation with a model of about 440MB is a notable achievement. However, actual translation quality and processing speed need to be verified on each device, and Korean translation accuracy in particular requires careful additional evaluation.