Xiaomi MiMo-V2.5-Pro-UltraSpeed tops 1,000 TPS

While Xiaomi's 1,000+ TPS on a trillion-parameter model is a stellar engineering feat, claims that this is the first useful speculative decoding method deployed on a quasi-frontier model are overstated.

* Speeds exceeding 1,000 TPS on commodity 8-GPU nodes demonstrate that massive models can be served efficiently without specialized supercomputers.

* DFlash's block-level masked parallel prediction offers a significant throughput improvement over traditional autoregressive draft-then-verify loops.

* Selective quantization (MXFP4 for experts, FP8 for attention) strikes a critical balance between reducing memory bandwidth bottlenecks and maintaining reasoning capabilities.

* The release of Hugging Face checkpoints invites open-source validation of these performance claims in real-world scenarios.

* Veteran AI practitioners note that multi-token prediction and speculative decoding have been utilized in production for nearly two years, making the "first useful" claim historically inaccurate.