Tencent Hunyuan open-sources Hy-SimulMT simultaneous interpretation model
Tencent Hunyuan released and open-sourced Hy-SimulMT on Oct 26, a simultaneous interpretation model supporting bidirectional Chinese and English speech-to-speech translation. Audio can be streamed in continuously and the model translates while listening, emitting text and speech together, with real-time timbre cloning that preserves the speaker's vocal character and intonation. On the open long-form RealSI benchmark, the official StreamLAAL figure is about 3 seconds, the lowest among the compared systems, against a 4-to-5-second average lag for professional human interpreters. Scored with GPT-6.1-Sol, the model reaches 90.3 on FLEURS, 75.5 on CoVoST2, 81.2 on ACL6060 and 84.7 on RealSI-Sentence, the last being the highest of every listed system against 82.2 for Gemini 3.5 LiveTranslate. In a real-scenario comparison with six CATTI-certified human interpreters across 10 domains and 18 recordings of about six minutes, the model matched the human quality level with lower latency. It is a 30B-A3B mixture-of-experts model with FP8 and GGUF builds plus an Interleave variant that outputs the source transcript, released under Apache 2.0 with the vLLM-based SimulInfer serving stack and the SimulTrain training tool. Weights are on Hugging Face and ModelScope, and the model ships first in the Hy Translate app at no charge.