Tencent Hunyuan has released Hy-SimulMT, a simultaneous translation model for spoken Chinese and English. It processes audio as it arrives and starts producing translated text without waiting for the speaker to finish a sentence. Alongside the model weights, the release includes tools for deployment and further training.
An external service handles speech output
The translation model itself produces text. For spoken output, the server passes that text to a separately configured external text-to-speech service, which also handles voice cloning rather than Hy-SimulMT itself. In an English-to-Chinese demonstration, Tencent showed translated speech retaining the distinct voices of three speakers.
Streaming output can include corrections when later context changes the meaning. Wang Weiwei, a vice dean at an interpreting school, described the system as starting to speak early to keep delays low, then correcting itself afterward.
Six variants offer two text output modes
Six variants are available on Hugging Face, using a mixture-of-experts architecture with 30 billion parameters in total and approximately 3 billion active per token. The base Hy-SimulMT-30B-A3B produces translated text, while Hy-SimulMT-30B-A3B-Interleave alternates between a transcript of the original speech and its translation.
Both output modes are available with full-precision weights for vLLM. Their FP8 variants use less memory and are intended for deployment on a single GPU.
Quantized GGUF packages are provided for llama.cpp, but official support has not yet been merged into the project’s main branch. Using these packages therefore requires the project’s current integration procedure or a compatible build.
Microphone input, uploaded files and custom API clients
The release includes SimulInfer, an inference service with a web interface and WebSocket communication. The interface accepts microphone input and uploaded audio files. Developers can use the API to build streaming translation clients, and the service offers optional text-to-speech integration.
SimulTrain provides scripts for continuing training from an existing model checkpoint. The released model supports Chinese and English only.
Tencent reports low latency, but trails on CoVoST2
In Tencent’s evaluation on long-form RealSI recordings, Hy-SimulMT recorded approximately three seconds of lag under StreamLAAL, a metric where lower values indicate less streaming translation lag. The company reports that this was the lowest latency among the systems compared.
For the same long-form recordings, Tencent reports higher GEMBA scores than every open-source model included in the comparison and four of the five commercial APIs. GEMBA assigns higher scores to better-quality translations.
Tencent also reports the best results among the open-source systems it compared on FLEURS, ACL6060 and RealSI-Sentence. On CoVoST2, however, the company’s results put Hy-SimulMT behind Qwen3.8-LiveTranslate.
A comparison with six experienced interpreters
Tencent also compared the model with six experienced interpreters using 18 Chinese and English recordings, each approximately six minutes long. Three interpreters held CATTI Level 1 certification, while the other three held CATTI Level 2 certification. The company says Hy-SimulMT delivered comparable quality in this evaluation and usually responded faster.
