Tencent Challenges Google with Its Offline Translation AI

Tencent recently launched an open-source AI translation model that could redefine industry standards. Named Hy-MT1.5-1.8B-1.25bit, this model stands out for its ability to operate entirely offline on smartphones while outperforming Google Translate, according to Tencent.

A Compact and Powerful Model

The model covers 33 languages, including widely spoken languages such as German, English, Chinese, Japanese, and French, as well as less common languages like Tibetan and Mongolian. It also supports five dialects and offers 1,056 translation directions. What makes this model particularly impressive is its size: it weighs only 440 MB. Thanks to an advanced compression technique, it achieves this size by using only 1.25 bits per parameter, reducing its initial size from 3.3 GB to a quarter of that.

Top-Notch Performance

Despite its compactness, Tencent's model competes with much larger solutions, such as Qwen3-32B, on standard benchmarks. It is about 25% smaller and 10% faster than models using 1.67 bits compression, while maintaining translation quality comparable to models weighing several hundred gigabytes.

Accessibility and Demonstration

To allow users to test this technology, Tencent has developed a demonstration application for Android, available for APK download. This application enables translation of words in any app without requiring an internet connection. The model has already proven its effectiveness by winning 30 first places in international machine translation competitions.

A Trend Towards Local AI

Tencent is not alone in this race to optimize AI models for offline use. Google is also working on a similar model, named Gemma 4, which operates locally on smartphones. This trend could transform how users access translation services, making these tools more accessible and reliable, even without an internet connection.