Open-source translation model series supporting 33 languages. Hunyuan-MT-7B won first place in 30 of 31 categories at WMT25. Hy-MT2 (May 21, 2026) extends the family with 1.8B / 7B / 30B-A3B variants — the 7B and 30B-A3B reportedly outperform DeepSeek-V4-Pro and Kimi K2.6 on translation, while the 1.8B quantizes to 440MB at 1.25-bit. Ships with IFMTBench, a companion instruction-following translation benchmark.

Hy-SimulMT (October 26, 2026, Apache 2.0) moves the line from text to live speech: bidirectional Chinese–English simultaneous speech-to-speech translation that translates while the speaker is still talking and clones the speaker's voice in real time. The model is a 30B-A3B MoE in Tencent's HY V3 architecture (48 layers, 128 experts, 8 active) with speech input. Speech output relies on a separate TTS service. On the long-form RealSI benchmark Tencent reports a streaming lag (StreamLAAL) of about 3 seconds, the lowest in its comparison, with a GEMBA quality score above the compared open models and four of five commercial APIs, and 84.7 on RealSI-Sentence. A small test on 18 new six-minute recordings found it on par with CATTI-certified human interpreters while usually responding sooner, a result Tencent says applies only to that sample. The release includes the SimulInfer serving stack and SimulTrain training tools.

Model Details

Architecture DENSE
Parameters 30B
Active params 3B

Variants

Name Parameters Notes
Hunyuan-MT-7B 7B WMT25 champion (Sep 2025)
Hunyuan-MT-Chimera-7B 7B Ensemble learning architecture
HY-MT1.5-7B 7B Released Dec 2025, adds terminology and contextual translation
Hy-MT2-1.8B 1.8B Released May 21 2026; 1.25-bit quantization fits in 440MB
Hy-MT2-7B 7B Released May 21 2026; self-reported beats DeepSeek-V4-Pro and Kimi K2.6 on translation
Hy-MT2-30B-A3B 30B Released May 21 2026; MoE variant
Hy-SimulMT-30B-A3B 30B Released Oct 26 2026; simultaneous zh<->en speech-to-speech translation with voice cloning; StreamLAAL ~3 s on RealSI; FP8, GGUF and Interleave builds
nlpopen-weightefficiency