Tencent open-sources Hy-MT2 translation models, launches Hy Translate app

Tencent's Hunyuan team is open-sourcing its Hy-MT2 family of multilingual translation models, and it's also launching a mini-program called Tencent Hy Translate for users to try out. The models come in three sizes: 1.8 billion parameters, 7 billion, and a 30-billion-parameter mixture-of-experts version (30B-A3B). They support 33 languages plus five Chinese dialects, and the weights and code are up on Hugging Face, ModelScope, and GitHub. The models are also already adapted for hardware from ARM, Qualcomm, Intel, and Chinese chip makers Muxi and Tianshu Zhixin.

But the real star here is on-device performance. Using Hunyuan's AngelSlim/Sherry framework, the 1.8B model is quantized down to a ridiculous 1.25 bits, compressing it to just 440MB. On an Apple A15 chip, inference is 1.5 times faster than the previous 4-bit quantized version, making it fully capable of offline, real-time translation on a mainstream smartphone. The larger 7B and 30B-A3B versions, meanwhile, beat DeepSeek-V4-Pro on the FLORES-200 benchmark and hit about 96 to 99 percent of Gemini 3.1 Pro's performance in real-world and domain-specific tests. Tencent also open-sourced IFMTBench, a benchmark for evaluating how well models follow translation instructions.

The Tencent Hy Translate mini-program is live now, with features like voice input, style presets, and customizable commands. Full iOS and Android apps with offline inference support are coming soon.