Hy-SimulMT

Hy-SimulMT is a real-time simultaneous translation model released and open-sourced by Tencent Hunyuan on October 26, 2026. It performs bidirectional Chinese↔English speech-to-speech translation: it continuously ingests audio and translates while the speaker is still talking, producing both translated text and speech, and uses real-time voice cloning to keep the speaker's vocal characteristics.

The core model, Hy-SimulMT-30B-A3B (a ~30B-parameter MoE with ~3B active), is on Hugging Face and ModelScope, with the model card listing Apache-2.0. The GitHub repo Tencent-Hunyuan/Hy-SimulMT ships the SimulInfer inference service and SimulTrain training tools, and you can also try it in the Hy Translate app.

@HuggingPapers summed it up on X: "Tencent just released Hy-SimulMT on Hugging Face — A real-time simultaneous translation model with ~3s latency and voice cloning."

Key features

• Translate while listening: built for meetings, speeches and interviews, balancing waiting for context against speaking early
• Low latency: on the RealSI long-form benchmark its StreamLAAL (streaming lag) is about 3 seconds, the lowest of the systems compared (StreamLAAL is not the same as time to first speech)
• Quality: its GEMBA score beats the compared open-source models and four of five commercial APIs
• Voice cloning: target-language speech keeps each speaker's voice, which makes multi-speaker conversations easier to follow (speech output needs a configured TTS service)

Short-form benchmarks (self-reported, higher is better)

Model / API	FLEURS	CoVoST2	ACL6060	RealSI-Sentence
Hy-SimulMT	90.3	75.5	81.2	84.7
Qwen3.8-LiveTranslate	88.4	82.1	80.5	84.4
Gemini3.5-LiveTranslate	92.4	74.4	82.6	82.2
Seed-LiveInterpret2.0	85.8	67.7	74.8	80.0
GPT-Realtime-Translate	77.7	50.8	76.4	76.7

The team also ran a small study on 18 new Chinese and English recordings from September 2026 onward (about 6 minutes each, 10 domains) against six interpreters holding CATTI Level 1 or Level 2 certificates. It reports quality on par with expert human interpreters and often lower latency, while stressing that the findings apply only to that sample and setup.

Model variants

Model	Description
Hy-SimulMT-30B-A3B	translation-only output
Hy-SimulMT-30B-A3B-Interleave	source text plus translation
-FP8 / -GGUF versions of both	FP8 quantization / GGUF format

Deploy with vLLM

You need Linux, Python 3.12+, Git and an NVIDIA GPU with enough memory. Existing environments need vLLM >= 0.31 with HYV3ForCausalLM support plus simulst.

git clone https://github.com/Tencent-Hunyuan/Hy-SimulMT.git Hy-SimulMT
cd Hy-SimulMT/SimulInfer
bash deploy.sh --install-server --model tencent/Hy-SimulMT-30B-A3B-Interleave

When you see Ready: http://127.0.0.1:8080/, open the web page, choose a direction and upload audio or start the microphone (remote microphone access needs HTTPS). From the command line:

curl -N http://127.0.0.1:8080/translate/stream \
  -F direction=en-zh -F [email protected]

• Existing model API: bash deploy.sh --backend-url http://model-host:8800/v1
• Multi-GPU: --gpus 0,1 --tensor-parallel-size 2
• llama.cpp: GGUF support is not yet merged upstream; the team recommends waiting for a release that supports it
• Further training: use SimulTrain in the repo

Who it's for

• Meetings, live streams and interviews that need real-time Chinese–English interpretation
• Teams that want to self-host simultaneous translation on their own GPUs

The team notes that professional interpreters remain important where strict accountability or high-stakes decisions are involved. Contact: [email protected]