Xiaomi Achieves High TPS with MiMo V2.5; DFLash Model Released with Open-Source Promise

Core summary

One sentence to understand this update

Xiaomi is reportedly serving MiMo V2.5 at 1000-3000 tokens per second using DFlash & Persistent kernel, with the DFLash model now available and an open-source release promised soon.

Impact & opportunity

What this could mean

This demonstrates advanced inference optimization techniques for large models, offering inspiration for builders seeking high-performance local deployments.