Xiaomi's MiMo-V2.5 cuts attention compute to 10-layer GQA level

After permanently cutting API prices on its MiMo-V2.5 series to match DeepSeek, Xiaomi's large model team lead Luo Fuli took to X to explain how it stays profitable.

Luo revealed that the high-load inference engine still breaks even at the new prices. The savings come from a hybrid attention architecture and hierarchical KV cache optimization.

For the design goal of a 99% reduction in cache hit costs, Xiaomi's inference framework implements hierarchical KV cache optimization for sliding window attention (SWA). Production tests show the optimization boosts cached token capacity by 5x and cuts cache costs by 80%. Combined with cache read overlap between global attention modules, the system further crushes actual cache hit overhead.

As for the 60% to 80% reduction in base input and output costs, Luo credits the model's 1:7 inter-layer sparsity ratio — one global attention (GA) layer for every seven SWA layers. During long-text prefill, the 60 SWA layers only compute local sliding windows. That means the 70-layer MiMo-V2.5-Pro model's overall attention compute is equivalent to a traditional 10-layer global GQA model. This ultralow compute load slashes raw inference cost, leaving Xiaomi a 2x to 3x profit margin before the price cut. So the price drop reflects structural cost reduction, not loss-leading competition.

Luo said low-cost inference services help drive on-device AI demand. She advised large model companies to avoid mindless price wars and instead use algorithm-inference co-design to keep actual runtime costs below breakeven.