Tencent Open-Sources 440MB Translation Model That Runs Entirely Offline on Phones

A Pocket-Sized Polyglot

Tencent has open-sourced an ultra-compressed translation model small enough to run entirely offline on a smartphone, releasing it ahead of China's May Day holiday travel season. The model, dubbed Hy-MT1.5-1.8B-1.25bit, squeezes a 1.8-billion-parameter translation engine supporting 33 languages into just 440MB, roughly one-eighth of its original 3.3GB size. Once downloaded, the model requires no internet connection, no subscription, and no cloud processing, with all translation handled locally on the device itself.

The Compression Breakthrough

The dramatic size reduction relies on Sherry, a ternary quantisation framework that has been accepted at the ACL 2026 conference. Sherry enforces a so-called 3:4 fine-grained sparsity pattern: for every four model parameters, three are stored at 1-bit precision and one is set to zero, yielding an average of just 1.25 bits per parameter. Paired with a custom inference kernel optimised for mobile CPU instruction sets, the approach allows the model to run smoothly in the background even on memory-constrained devices. Tencent also released a 2-bit variant at 574MB that preserves near-lossless translation quality relative to the full-precision model.

Performance Punching Above Its Weight

The underlying Hy-MT1.5 model family, an upgrade of Tencent's WMT25 championship translation system, natively supports 33 languages, five Chinese dialects and minority languages, and 1,056 translation directions. Tencent claims the 1.8-billion-parameter model achieves translation quality comparable to commercial APIs and models with 235 billion parameters, and surpasses Google Translate on standard benchmarks. That is a remarkable claim for a model that fits comfortably alongside a few photos in a phone's storage.

Privacy by Design

Alongside the model weights, Tencent released a demo app featuring a background word-capture mode that lets users get translations whilst reading emails or browsing the web. The app processes everything on-device and collects no personal data, according to Tencent. The demo is currently available for Android through ModelScope, with iOS support not yet available.

A Growing Open-Source Lineup

The release adds to Tencent's broader Hunyuan translation lineup, which includes the larger HY-MT1.5-7B model for cloud deployment. The full model weights and demo app are open-sourced, making the 440MB offline translator freely available to developers and end users alike, a notable contribution to the growing ecosystem of edge-deployable AI.