Xiaomi went to optimize it's Mimo V2.5-Pro to squeeze the max out of regular GPUs, and not betting on specialized hardware like Groq or Cerebras. They combined: - FP4 quantization with QAT - DFlash speculative decoding - TileRT latency optimized kernels In close collaboration with the TileRT team they achieved 1000+ t/s on an 8-GPU cluster using this approach. It's available on their API at 3x the price of the normal API - once you have been granted access. Read Xiaomi's blog post her...