The model is a mixture-of-experts system with 30 billion total parameters and about 3 billion active parameters.
Hy-SimulMT is designed for bidirectional Chinese-English simultaneous speech translation: it takes a continuous audio stream while a speaker is still talking and produces translated text. Tencent says a separately configured TTS service can produce translated speech with real-time voice cloning. The release includes BF16, FP8, and GGUF weights, an Interleave variant, the SimulInfer inference service, and SimulTrain training tools.
Tencent reports about three seconds of StreamLAAL lag on the RealSI long-form benchmark, plus a GEMBA score above the open-source systems and four of five commercial APIs it compared. Those are vendor-reported comparisons; the source does not establish independent benchmark results, production reliability, TTS quality, or upstream llama.cpp support. The model card says llama.cpp support has not yet been merged upstream.