Hy-SimulMT
Tencent Hunyuan open-sourced Hy-SimulMT, a Chinese-English simultaneous interpretation model that clones the speaker's voice and lags about 3 seconds on long-form speech.
Hunyuan's earlier open translation models worked on text, and Hy-SimulMT moves to live speech to speech interpretation in both directions. It is a 30B-A3B mixture-of-experts model that translates while the speaker is still talking and clones the voice in real time through a text-to-speech service. The release includes weights, FP8 and GGUF builds, an Interleave variant that also outputs the source transcript, the SimulInfer serving stack and the SimulTrain training tools.
- Date
- Tuesday 27 October 2026
- Lab
- Tencent
- Kind
- open-weights
- Access
- open weights
- Price
- Free in the Hy Translate app
Figures
| Measure | Value | Measured by |
|---|---|---|
| StreamLAAL latency on RealSI long-form | about 3 seconds lowest among the systems Tencent compared; not the same as time to first translated speech | company |
| FLEURS score | 90.3 short-form | company |
| CoVoST2 score | 75.5 short-form | company |
| ACL6060 score | 81.2 short-form | company |
| RealSI-Sentence score | 84.7 highest of all listed systems | company |
Pandaily reports an Apache 2.0 license on the Hugging Face card. Speech generation and voice cloning need a configured text-to-speech service, and llama.cpp support is not yet merged.
Sources
- pandaily.com/tencent-hunyuan-hy-simulmt-open-source-simultaneous-interpretation-voice-clon
- technode.com/2026/10/27/tencent-open-sources-ai-model-for-real-time-chinese-english-interp
This record was checked against its sources on 27 October 2026. How we check