Tencent Opens Hy-SimulMT Chinese-English Simultaneous Translation Model, with Streaming Inference and Training Tools

Tencent Hunyuan released Hy-SimulMT on October 26, publishing the model weights, the SimulInfer inference service, and the SimulTrain training tools. Designed for meetings, talks, and interviews, the system continuously receives audio and generates translations before the speaker has finished a full passage. This lets developers deploy and adapt their own streaming translation workflows. Official model card

The release includes versions that output only the translation and versions that output the original and translated text together in an Interleave format, as well as FP8 and GGUF formats. SimulInfer includes a web interface, audio file uploads, and real-time WebSocket support. Speech synthesis and voice cloning require an external TTS service, so downloading the translation weights alone is not enough to provide a complete bidirectional voice experience. Project documentation

For performance, the team reported a StreamLAAL of about three seconds on the long-audio RealSI test and said its GEMBA translation quality score exceeded those of the open-source models in the comparison and four of five commercial APIs. StreamLAAL measures how far streaming translation lags behind the source content; it should not be interpreted directly as the wait time before translated audio first plays. Short-audio tests also showed mixed results: the model scored a listed-best 84.7 on RealSI-Sentence, but did not lead on every dataset. These are all official evaluation results. Evaluation methods and results

For deployment, the official vLLM path requires Linux, Python 3.12 or later, and vLLM 0.31 or later with support for HYV3ForCausalLM. The existing model service can be started separately from the web service, making it easier to connect to an existing inference environment. GGUF users should note that the project explicitly says the related llama.cpp support has not yet been merged upstream. The proposed support covers command-line inference; server support remains future work. Deployment and compatibility notes

Engineering teams should next measure translation quality, time to first translated audio, and long-duration streaming stability separately, then assess end-to-end latency with TTS added. The official release also reports a small-scale comparison using 18 recordings and six interpreters, but limits its conclusions to that sample and setup; it does not cover a comprehensive assessment of speech quality or voice similarity. Testing limitations