Tencent Unveils Multilingual Translation Model Hy-MT2

Tencent has released the Hy-MT2 series, a 'fast-thinking' multilingual translation model optimized for complex real-world scenarios. The models come in three sizes—1.8B, 7B, and 30B-A3B (MoE)—and support translation and multilingual instruction-following across a total of 33 languages.

Key Features and Performance:

• 1.8B model: Through AngelSlim 1.25-bit extreme quantization, storage size is reduced to 440MB and inference speed is improved by 1.5x, optimizing it for on-device deployment. In terms of performance, it surpasses major commercial APIs from Microsoft and Doubao.
• 7B and 30B-A3B models: In 'fast-thinking' mode, they outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6.

Tencent also open-sourced IFMTBench, a benchmark for evaluating translation instruction-following capability. In addition, it has partnered with WMT26's video subtitle translation task to support the related competition.