Xiaomi hits 1,000 tokens per second, running massive AI on standard hardware

Xiaomi announced the MiMo-V2.5-Pro-UltraSpeed, the first 1-trillion-parameter AI model to exceed 1,000 tokens per second. Performance peaked at nearly 1,200 tokens per second during testing.

Engineers achieved these speeds on a standard server using eight commodity GPUs. The system utilizes FP4 quantization and DFlash speculative decoding to reduce the cost of deploying high-performance models.

Xiaomi will host a limited, application-based API trial from June 9 to June 23. The accelerated service costs three times the standard rate while delivering ten times the generation speed.