Xiaomi and TileRT Achieve 1200 Tokens Per Second on 1-Trillion-Parameter Model

Xiaomi and TileRT achieve 1200 tokens per second on a 1-trillion-parameter model using MiMo-V2.5-Pro-UltraSpeed. The result runs on a standard 8-GPU commodity node with FP4 quantization, DFlash speculative decoding, and TileRT system optimization. UltraSpeed boosts decode speed without reducing output quality through coordinated model and serving system design.