Tencent-Hunyuan has open-sourced Hy-SimulMT, a 30B-parameter mixture-of-experts model that simultaneously interprets between Chinese and English speech, cloning the speaker's voice in real time1,2. The release, published on October 27, includes model weights, inference software, and training tools under an Apache 2.0 license.
The headline metric is latency. Measured by StreamLAAL, a standard metric for streaming translation lag, Hy-SimulMT posts approximately 3 seconds on the long-form RealSI benchmark, the lowest among the systems Tencent compared. The project notes that StreamLAAL is not the same as the time until the first translated speech is heard. For context, IThome puts the typical lag of professional human interpreters at 4 to 5 seconds.
Quality scores and human comparison
On translation quality, Hy-SimulMT's GEMBA score on long-form speech beat every open model in the comparison and four of the five commercial APIs tested. Short-form benchmark scores were 90.3 on FLEURS, 75.5 on CoVoST2, 81.2 on ACL6060, and 84.7 on RealSI-Sentence, the last being the highest of every listed system.
Tencent also compared the model against human interpreters on 18 recordings of roughly six minutes each, recorded from September 2026 onward across 10 domains.
Wang Weiwei said the system speaks first and corrects itself afterward, much like a human interpreter, and that AI interpreting is moving from helping people understand toward usable communication in low-risk settings.
Serving stack and model formats
The release ships with SimulInfer, a serving stack built on vLLM with a web interface, a WebSocket API, and optional text-to-speech, as well as SimulTrain for further training on domain data. FP8 and GGUF builds are available, along with an Interleave variant that also outputs the source transcript. Speech generation and voice cloning require a configured text-to-speech service. Support in llama.cpp has not yet been merged.
Hy-SimulMT is live and free to use in the Hy Translate app.
ANALYSIS The 30B-A3B mixture-of-experts architecture keeps active parameters small relative to total model size, a design choice that bears on deployment cost for a latency-sensitive, streaming workload. Shipping a vLLM-based serving stack and GGUF weights lowers the integration bar for teams already running those toolchains.