AsianFin -- Xiaomi announced on Sunday the full open-sourcing of its MiDashengLM-7B large model, which the company claims has set new state-of-the-art (SOTA) benchmarks in audio understanding across 22 public evaluation datasets.
According to Xiaomi, MiDashengLM-7B achieves first-token inference latency (TTFT) at just one-fourth that of leading models in the industry, while delivering over 20 times greater data throughput efficiency on equivalent GPU memory. The model’s performance underscores its potential for high-speed, low-latency applications in multimodal AI scenarios.
Xiaomi said it is already working on further optimizing MiDashengLM’s computational efficiency, aiming for offline deployment on consumer-grade devices in future iterations.
The open-sourcing move marks Xiaomi’s latest effort to position itself at the forefront of AI-driven smart device ecosystems, as competition in edge AI and multimodal models heats up.
Explore more exclusive insights at nextfin.ai.
