Xiaomi’s MiMo‑V2.5‑Pro UltraSpeed delivers a 1‑trillion‑parameter model that generates over 1000 tokens per second on a standard 8‑GPU server by combining FP4 quantization, MoE architecture, DFlash decoding and TileRT’s custom execution engine, challenging the need for dedicated ASICs.
The article explains how Xiaomi’s MiMo‑V2.5‑Pro‑UltraSpeed mode achieves a record‑breaking 1000 tokens per second inference speed, why such ultra‑fast performance matters for real‑time AI applications, and the FP4 quantization, DFlash decoding and TileRT inference technologies that make it possible without sacrificing model quality.
Xiaomi’s MiMo‑V2.5‑Pro UltraSpeed mode breaks the 1000 tokens‑per‑second barrier for a 1‑trillion‑parameter model by combining FP4 expert‑only quantization, DFlash block‑masked speculative decoding, and TileRT’s ultra‑low‑latency GPU system, and the API is now available through a limited‑time trial.
On May 22, Zhipu launched the GLM‑5.1‑highspeed API, delivering 400 tokens per second—about 7× faster than the original model and twice as fast as Gemini 3.5 Flash—through a three‑layer optimization that rewrites the MoE inference path, introduces dynamic scheduling, and leverages TileRT’s AOT engine to cut latency while preserving full flagship capabilities.
