Xiaomi announces the release of the open-source large-scale voice model MiDashengLM-7B.

On August 4th, Xiaomi released and fully open-sourced the MiDashengLM-7B model. According to the introduction, the sound understanding performance of MiDashengLM-7B surpasses the best performance of multi-modal large models on 22 public evaluation sets. The first token latency of single sample inference is only 1/4 of the industry's advanced models, and the data throughput efficiency under the same amount of memory is more than 20 times that of the industry's advanced models. Building on the current version, Xiaomi has begun to further upgrade the computing efficiency of the MiDashengLM model, seeking to enable offline deployment on terminal devices.