On August 4th, Xiaomi released and fully open-sourced the MiDashengLM-7B model. The MiDashengLM-7B's sound understanding performance set new state-of-the-art (SOTA) scores for large multimodal models across 22 public evaluation datasets. Its latency to first token (TTFT) for single-sample inference is only one-quarter that of the industry's leading models, and its data throughput is over 20 times that of the industry's leading models using the same amount of video memory. Building on the current version, Xiaomi is working to further improve the computational efficiency of the MiDashengLM model and is pursuing offline deployment on mobile devices.