Tencent’s HunYuan team has unveiled an open-source translation model called Hy-MT1.5-1.8B-1.25bit, designed for lightweight offline use. According to the release, the model compresses 1.8 billion parameters into a package of just 440MB, making it suitable for smartphones and allowing it to run without an internet connection while continuing to operate in the background on mobile devices.
Support for 33 languages and 1,056 translation paths
The model supports 33 languages and 1,056 translation directions. Its coverage includes major languages such as Chinese, English, and French, as well as lower-resource languages including Tibetan and Mongolian. That breadth suggests potential use cases across everyday communication, multilingual content handling, travel, and education scenarios where offline access can be especially valuable.
Compression technology is the key differentiator
Tencent said the model relies on its Sherry compression technology, which has been recognized by ACL 2026. Despite its reduced size, the company claims the model outperformed Google Translate in comparative testing and delivered performance comparable to much larger systems. Tencent also released a 2-bit version, which it described as offering near-lossless accuracy.
Weights, code, and test builds are already available
For developers and researchers, Tencent has made the model weights, source code, and technical report available on Hugging Face and ModelScope. On the product side, an Android APK is already available for testing, while an iOS version is expected to follow. The release highlights a broader industry trend toward bringing advanced AI capabilities from the cloud onto local devices through aggressive compression and lower deployment requirements.