Xiaomi is now serving MiMo V2.5 at 1000-3000tps using DFlash & Persistent kernel. DFLash model is out, open-source release promised coming soon

Xiaomi has announced the release of MiMo V2.5, which operates at a throughput of 1000-3000 transactions per second (tps) utilizing DFlash and a Persistent kernel architecture. An open-source version of the DFlash model is expected to be available soon. This development is significant for practitioners as it enhances the performance capabilities of AI systems, particularly in real-time applications requiring high throughput.