Tencent Open‑Sources Hy‑MT: Offline Translation for 33 Languages Beats Google Translate

Model Overview

Hy‑MT1.5‑1.8B‑1.25bit is an open‑source neural machine translation model released by Tencent Mixed‑Model team. It has 1.8 billion parameters and is quantized to 1.25 bit precision, resulting in a model size of a few hundred megabytes that fits in mobile device storage.

Breaking the Offline Translation Trade‑off

Traditionally high translation quality required large models, large models limited language coverage, and broad coverage prevented offline execution. Hy‑MT1.5‑1.8B‑1.25bit simultaneously achieves:

Support for 33 language pairs, including Chinese, English, Japanese, Korean, French, German, Arabic and several low‑resource languages.

Fully offline inference after a one‑time download; no network connection is needed.

Execution on Snapdragon or Dimensity mobile SoCs without cloud assistance.

The key technical enabler is a 1.25‑bit quantization pipeline combined with a specialized training strategy. Conventional quantization typically stops at 4 bit or 8 bit because lower precision degrades fluency and accuracy. The authors’ algorithm preserves translation quality at the extreme 1.25‑bit compression rate, allowing the 1.8 B‑parameter model to run smoothly on‑device.

Performance Compared with Google Translate

Internal benchmark tests show Hy‑MT1.5‑1.8B‑1.25bit surpasses Google Translate’s online version on medium‑ and low‑resource language pairs and on colloquial expressions. The evaluation notes that Google Translate still leads in overall language breadth and ecosystem integration, but the offline model matches or exceeds the quality of the online service in the tested scenarios.

Implications of Open‑Source Release

By publishing the model, phone manufacturers and application developers can embed it directly into operating systems, instant‑messaging apps, or vehicle infotainment systems, turning translation from a fallback feature into a default, always‑available service. The model’s offline nature also ensures that user data never leaves the device, addressing privacy concerns in business or medical translation contexts.

Limitations

The 1.8 B‑parameter size is not the smallest possible for edge deployment.

Extreme 1.25‑bit quantization imposes some flexibility constraints on model adaptation.