Tencent Hy-MT2-1.8B

Hy-MT2-1.8B is the smallest of the three models in Tencent's second-generation translation family, published on 21 May 2026. It is a dense model of 2.04 billion parameters — no mixture of experts, no routing — and it supports the same 33 languages and the same instruction-driven translation features as its 30-billion-parameter sibling: enforced glossaries, named styles, exact delimiter preservation, and translating only the human-readable values inside JSON or XML. The reason to look at it is size. Using AngelSlim, Tencent's own compression toolkit, the model is quantised to 1.25 bits per weight, which brings storage down to 440 MB and, by the company's measurement, makes inference about 1.5 times faster. At that size a translator between 33 languages fits on a phone with room to spare, which is a different product from a translation API: no round trip, no per-request cost, and no text leaving the device. The 1.25-bit build and a plain GGUF build are both published. Downloads suggest readers have noticed. Over the 30 days to 31 August 2026 the llama.cpp build was pulled 213,580 times against 24,302 for the base weights — a ratio that says most of the interest is in running this locally rather than fine-tuning it. Note that Hugging Face's counter only ever covers a rolling 30 days, so these are current rates, not lifetime totals. Tencent's own comparison claims the 1.8B beats mainstream commercial translation APIs, naming Microsoft and Doubao, taken overall. That is a vendor chart and should be read as one; what can be checked independently is that the weights are downloadable and that the licence in the repository is the plain Apache 2.0 text, with none of the territorial exclusion that the company's HunyuanOCR models carry against the EU, the UK and South Korea.

▸Videos

No videos yet.

▸Reviews