Jiepao Xingchen has launched its advanced automatic speech recognition model, StepAudio 2.5 ASR, featuring Multi-Token Prediction (MTP) technology. This innovation accelerates inference speed and utilizes a 32K context window, allowing seamless transcription of 30-minute audio without slicing. The model's ASR+MTP-5 architecture boosts inference throughput by 400%, reduces latency by 60%, and cuts costs by 80%, achieving a peak rate of 500 tokens per second. Tests show improved accuracy and lower word error rates compared to competitors.