Key Info

Tencent has open-sourced Hy-SimulMT, a simultaneous Chinese-English speech translation model that runs on your own GPUs and streams translation as people speak.

Highlights

  • Holds latency to about 3 seconds on the open RealSI long-form benchmark and scores 84.7 on RealSI-Sentence.
  • Six checkpoints released under Apache 2.0, including FP8 and GGUF versions.
  • Repo ships an inference service plus training tools.
  • Paired with a TTS service, it can respond in the speaker's own voice in Chinese or English.