Tencent open-sources Hy-MT2, shrinking a translation LLM to 440MB

Hy-MT2 is Tencent's newly open-sourced multilingual translation model built specifically for app localization, subtitle translation, and edge deployments. It supports 33 languages across three distinct sizes: 1.8B, 7B, and a 30B-A3B MoE architecture.

Tiny footprint with tight formatting control

The standout feature here is just how easy it is to run on local hardware. Using AngelSlim 1.25-bit extreme quantization, the 1.8B model shrinks down to a featherweight 440MB, meaning it can run comfortably on phones or edge devices. In terms of output quality, Tencent notes that this compact 1.8B version generally outperforms commercial translation APIs from Microsoft and Doubao.

Unlike traditional machine translation engines that routinely break code and markup, Hy-MT2 doubles down on steerability and instruction following. It handles terminology consistency, matches target writing styles, preserves custom delimiters, and translates only user-facing values inside structured formats like JSON while keeping the syntax intact.

Why it matters

If you are localizing apps, games, or video subtitles, Hy-MT2 is well worth a spin. Squeezing a capable translation engine into just 440MB lets developers skip recurring cloud API bills and run tasks completely offline without data privacy concerns.

We think Tencent prioritized the right problem here. The hardest part of production translation usually isn't grammar—it is keeping structured data unbroken and terminology aligned. Hy-MT2 tackles that practical developer pain point head-on.