Tencent Hunyuan Open-Sources Hy-SimulMT, a ~3s-Latency Real-Time Speech Translator

On October 26, Tencent's Hunyuan team officially open-sourced Hy-SimulMT, a real-time Chinese-English simultaneous interpretation model, now available on Hugging Face, ModelScope, and GitHub (Tencent-Hunyuan/Hy-SimulMT) for anyone to download and use. It's a rare move for a major Chinese tech company to open-source not just a simultaneous interpretation model but its full training pipeline as well—well worth attention.

Confirmed

• The flagship version is Hy-SimulMT-30B-A3B, built on a MoE architecture and offered in multiple formats including FP8, GGUF, and Interleaved (the GGUF version makes local deployment easier).
• Supports bidirectional Chinese-English speech-to-speech simultaneous interpretation: the model continuously receives an audio stream and progressively produces translation text and speech while the speaker is still talking.
• Capable of real-time voice cloning, so the translated speech retains the speaker's voice.
• Streaming latency (StreamLAAL) is around 3 seconds; according to @aigclink, some tests are already close to professional interpreter level.
• The GitHub repo includes the SimulInfer inference service and SimulTrain training pipeline, with both Chinese and English READMEs covering the full inference and training workflow.
• Target use cases include real-time communication scenarios such as meetings, speeches, and interviews.

Why it matters

• Traditional machine translation waits until the source sentence is finished before translating, whereas Hy-SimulMT's streaming approach is much closer to real-world meeting scenarios—its roughly 3-second latency is practically competitive for simultaneous interpretation.
• Direct speech-to-speech translation combined with real-time voice cloning makes the translated speech sound more natural, avoiding a robotic tone that disrupts the flow of conversation.
• What's open-sourced isn't just the weights but the complete training pipeline, giving the community a foundation to reproduce and improve simultaneous interpretation models.

2026-10-26 ~ 2026-10-26 · 7 related posts

• Tencent Hunyuan open-sources Hy-SimulMT-30B: bidirectional CN-EN live speech translation with voice cloning — jacek2023 · 2026-10-26
• Tencent open-sources Hy-SimulMT: real-time Chinese-English simultaneous translation with ~3s latency and voice cloning — TencentHunyuan · 2026-10-26
• Tencent open-sources Hy-SimulMT: real-time speech-to-speech translation with ~3s latency — aigclink · 2026-10-26
• [source] Tencent's Hy-SimulMT lands on GitHub with SimulInfer inference and SimulTrain pipeline — aigclink · 2026-10-26
• Tencent open-sources Hy-SimulMT: 30B/A3B realtime translation with 3s lag — AdinaYakup · 2026-10-26
• [source] Tencent's Hy-SimulMT lands on Hugging Face in six variants — AdinaYakup · 2026-10-26

1 near-duplicate retellings: TencentHunyuan