Xiaomi announces open source sound understanding big model MiDasheng LM-7B

On August 4th, Xiaomi released and fully opened sourced the MiDasheng LM-7B model. It is reported that MiDashengLM-7B has achieved the best performance in multimodal large models (SOTA) on 22 public evaluation sets in terms of sound understanding performance. The first token delay (TTFT) of single sample inference is only 1/4 of the industry's advanced models, and the data throughput efficiency under the same VRAM is more than 20 times that of the industry's advanced models. On the basis of the current version, Xiaomi has started to further upgrade the computational efficiency of the MiDashengLM model and seek offline deployment on terminal devices.