Xiaomi MiMo open-sources a 1T model running over 1,000 tps on 8 GPUs

Xiaomi MiMo and the TileRT team say a 1T model exceeds 1,000 tps on a single standard 8-GPU node using general-purpose GPUs.

The speedup comes from FP4 quantization and DFlash, a block-masked parallel speculative decoding method that accepts more tokens per verification, with TileRT tailoring its compiler and kernels to these techniques.

Open weights for the FP4 + DFlash checkpoint are available on Hugging Face.