Hy‑MT2‑30B‑A3B: Tencent's Massive Multilingual Translation Engine Hits the Spotlight

Tencent's Hy‑MT2‑30B‑A3B is the flagship member of the Hy‑MT2 family, a set of "fast‑thinking" multilingual translation models released in May 2026. Built on the Transformers library and released as a Mixture‑of‑Experts (MoE) 30‑billion‑parameter model, it supports instruction‑following translation across 33 languages, ranging from Chinese and English to less‑common languages such as Tibetan, Kazakh, and Uyghur. The model is tagged for translation tasks and is distributed in safetensors format, with a license under Apache‑2.0.

The README highlights that Hy‑MT2‑30B‑A3B, together with its 1.8B and 7B variants, achieves strong results on both general and domain‑specific benchmarks. It outperforms open‑source rivals like DeepSeek‑V4‑Pro and Kimi K2.6 in "fast‑thinking" mode and even surpasses commercial APIs from Microsoft and Doubao on many metrics. The release also bundles IFMTBench, a benchmark for evaluating translation instruction‑following capabilities, and provides a "Hy‑MT2‑Translator Skill" for easy integration into AI‑assistant platforms.

Deployment guidance is provided for several runtimes, including standard Transformers pipelines, vLLM, and SGLang, with recommended inference parameters (e.g., temperature 0.7, top‑p 1.0, max tokens 4096). For on‑device scenarios, the smaller 1.8B model can be quantized down to 1.25‑bit using AngelSlim, shrinking storage to 440 MB and boosting speed by 1.5×. The model is also part of Tencent's partnership with WMT26 for the Video Subtitle Translation Task, inviting the community to compete using Hy‑MT2 models.

Overall, Hy‑MT2‑30B‑A3B offers a powerful, open‑source solution for large‑scale, multilingual translation with instruction‑level control, making it attractive for developers building translation services, subtitle pipelines, or multilingual AI assistants.