Tencent's Hunyuan team open-sourced Hy-MT2, a translation model family in three sizes (1.8B, 7B, and a 30B-A3B mixture-of-experts) on Thursday. The models cover 33 languages plus five dialects and minority languages including Cantonese, Tibetan, Kazakh, Mongolian, and Uyghur. Weights are live on Hugging Face.
The interesting part is the small one. Tencent compressed the 1.8B model to 1.25 bits with its AngelSlim toolkit, dropping storage to 440MB and claiming a 1.5x inference speedup. That puts a usable translator inside the memory footprint of a mid-range phone, no API calls required.
On benchmarks, Tencent says the 7B and 30B-A3B models beat open-source rivals DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the 1.8B version surpasses Microsoft Translator and Doubao overall. Numbers are self-reported. The accompanying technical report claims the 7B and 30B-A3B variants reach 97.9% and 98.6% of Gemini 3.1 Pro Think respectively, though independent testing hasn't landed yet.
Tencent also released IFMTBench, a benchmark for measuring how well translation models follow user instructions like terminology constraints, style requirements, and structured data preservation. The team is sponsoring the WMT26 video subtitle translation track.
Bottom Line
The quantized 1.8B Hy-MT2 fits in 440MB, small enough to ship inside a mobile app without an API dependency.
Quick Facts
- Three model sizes: 1.8B, 7B, and 30B-A3B (MoE)
- 1.8B model compressed to 440MB via 1.25-bit AngelSlim quantization
- 1.5x inference speedup on quantized 1.8B, company-reported
- 33 languages plus 5 dialects and minority languages supported
- Released May 21, 2026 on Hugging Face and ModelScope
