China's Xiaomi MiMo Is Now 15X Faster Than ChatGPT and Claude (decrypt.co)

Xiaomi, in partnership with TileRT, has achieved a remarkable milestone with its MiMo-V2.5-Pro-UltraSpeed model, now delivering over 1,000 tokens per second on a standard 8-GPU commodity node, marking a significant advancement in AI inference speed. This performance, which peaks near 1,200 tokens per second, dramatically outpaces current competitors like ChatGPT and Claude, whose speeds range from 68 to 192 tokens per second. By leveraging innovative techniques such as FP4 quantization—reducing the precision of expert layers to 4 bits without sacrificing quality—and DFlash speculative decoding, which enables faster token processing, Xiaomi has demonstrated that high-speed AI performance can be achieved without relying on custom hardware. This breakthrough has profound implications for the AI/ML community as it potentially lowers the barrier for deploying high-speed inference in production environments. The efficiency gains allow for complex real-time applications, such as fraud detection and trading signals, to operate within strict latency constraints—something that slower models cannot achieve. For developers, a limited API trial will be available from June 9 to June 23, offering this advanced capability at a cost of three times the standard MiMo rates, reflecting its superior performance. With the FP4-DFlash checkpoint already open-sourced on Hugging Face, the community is invited to test and explore the capabilities of this cutting-edge technology.