MiMo v2.5 inference optimization improves hybrid SWA efficiency

Xiaomi published a blog post detailing inference optimizations for its MiMo v2.5 model. The techniques focus on improving hybrid Sliding Window Attention (SWA) efficiency for faster inference. No specific performance numbers were provided in the title or snippet.