Tencent Hunyuan · 2026-05-21 · notable
Tencent Hy-MT2 — Open-Weight Multilingual Translation Family (1.8B / 7B / 30B-A3B MoE) Covers 33 Languages; 1.8B Quantizes to 440MB at 1.25-bit
Tencent open-sourced a three-size translation model family spanning 33 languages. The 7B and 30B-A3B beat DeepSeek-V4-Pro and Kimi K2.6 on translation; the 1.8B shrinks to 440MB at 1.25-bit for on-device use.
An open-weight family of Tencent translation models spanning 1.8B to a 30B MoE, covering 33 languages.
Key specs
| Parameters | 1.8B / 7B / 30B-A3B MoE |
|---|---|
| Languages | 33 |
| 1.8 b quantized | 440MB at 1.25-bit |
What is it?
Hy-MT2 is a family of multilingual machine-translation models from Tencent's Hunyuan team, released in 1.8B, 7B, and 30B-A3B (mixture-of-experts) sizes. All three handle translation across 33 languages and follow translation instructions — terminology, style, and personalization — written in multiple languages.
How does it work?
The models use a fast-thinking decoding mode tuned for translation rather than open-ended chat. The 30B variant is a sparse MoE that activates roughly 3B parameters per token. The 1.8B model is compressed with AngelSlim 1.25-bit quantization to about 440MB for on-device deployment, with around 1.5x faster inference.
Why does it matter?
Tencent reports the 7B and 30B models beat open-source systems like DeepSeek-V4-Pro and Kimi K2.6 on translation in fast-thinking mode, while the 1.8B model surpasses commercial APIs from Microsoft and Doubao — giving developers open weights they can run locally instead of paying per-call translation APIs.
Who is it for?
developers and teams building multilingual translation
Try it
tencent/Hy-MT2-30B-A3B on Hugging Face