Tencent Hunyuan open-sources 440MB offline translation model, claims Google Translate quality lead

腾讯混元最新开源：440M翻译模型手机离线就能用，翻译质量超谷歌

Tencent Hunyuan open-sourced Hy-MT1.5-1.8B-1.25bit, compressing a 1.8B translation model to 440MB. It supports 33 languages and 1,056 directions, with an Android demo running offline on Snapdragon 888 and 8GB RAM. The key detail is Sherry 1.25-bit quantization: 3 of every 4 weights use 1 bit and 1 is zeroed.

Why it matters: HKR-H/K/R all pass: the story has a strong offline-phone hook, concrete quantization details, and practitioner relevance around edge inference. It stays below P1 because this is a vertical translation model, not a major foundation-model release.