Luo Fuli reveals MiMo's cost reduction strategy: Pre-filled attention computation reduced to 10 layers of global GQA level.

According to Beating's monitoring, after the Xiaomi MiMo-V2.5 series of self-developed large-scale models implemented a permanent API price reduction, Luo Fuli, head of the Xiaomi large-scale model team, announced an algorithm cost reduction mechanism on the X platform. Luo Fuli revealed that after aligning the API price with DeepSeek, Xiaomi's high-load inference engine can still maintain break-even. The cost reduction mainly comes from the hybrid attention architecture and hierarchical KV caching optimization. To achieve the design goal of reducing cache hit costs by 99%, Xiaomi's inference framework implemented hierarchical KV caching optimization for sliding window attention (SWA). Production testing shows that hierarchical optimization increases the cached token capacity by 5 times and reduces caching costs by 80%. Combined with cache read overlap technology between global attention modules, the system further reduces the actual overhead of cache hits. Luo Fuli attributes the 60% to 80% reduction in basic input and output costs to the 1:7 inter-layer sparsity ratio introduced in the model, meaning the ratio of global attention (GA) to sliding window attention (SWA) layers is 1:7. In the long text prefilling stage, the 60-layer SWA only calculates local sliding windows, making the overall attention computation of the 70-layer MiMo-V2.5-Pro model equivalent to that of a traditional 10-layer global GQA model. This ultra-low computational load reduces the original inference cost, allowing Xiaomi to reserve a 2 to 3 times profit margin before the price adjustment. Therefore, the price reduction reflects structural cost reduction, not loss-making competition. Luo Fuli stated that low-cost inference services are conducive to stimulating demand for intelligent terminals. Large model companies should avoid blind price wars and control actual operating costs below the break-even point through the underlying collaborative design of algorithms and inference systems.