Skip to main content
Back

Tencent's Hy-MT2 Squeezes 33-Language Translation Into 440MB on Your Phone

Tencent released Hy-MT2, a translation family spanning 1.8B to 30B-A3B MoE, with a 1.25-bit variant that fits in 440MB and runs offline.

  • Tencent released Hy-MT2, a translation model family in 1.8B, 7B, and 30B-A3B MoE sizes.
  • All variants support 33 languages including Tibetan, Uyghur, Mongolian, and Cantonese, under Apache 2.0.
  • 1.25-bit AngelSlim quantization shrinks the 1.8B model to 440 MB with a 1.5x speedup for on-device use.
  • 7B and 30B-A3B are claimed to beat DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode.
  • New IFMTBench benchmark evaluates translation instruction-following including terminology, style, and structured data.
  • Ships with transformers, vLLM, SGLang support, plus FP8 and GGUF variants for local inference.

Tencent's Hunyuan team just released Hy-MT2, a new generation of specialized translation models spanning three sizes and pushing hard on both quality and on-device deployment. The lineup includes a 1.8B dense model, a 7B dense model, and a 30B-A3B mixture-of-experts model, all released under Apache 2.0 and trained to translate across 33 languages.

The headline claim: the larger variants beat closed and open competitors in fast-thinking mode, while the tiny 1.8B version squeezes down to 440 MB through 1.25-bit quantization and still holds its own against commercial APIs. That combination of scale range and aggressive compression is what makes this release worth a closer look.

Three sizes, one translation stack

Hy-MT2 is pitched as a family of fast-thinking translators built for messy real-world inputs rather than clean benchmark sentences. According to the model card, the 7B and 30B-A3B variants outperform open-source models like DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the 1.8B model surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.

Evaluation covers four axes visualized in the release chart: general translation on WildMTBench and FLORES-200, competition-grade WMT25, Chinese minority language pairs, and radar plots for instruction following and domain-specific translation.