Luffy Decrypts MiMo Cost Reduction Strategy: Pre filled Attention Calculation Reduced to 10 Layers Global GQA Level According to Beating monitoring, after implementing a permanent API price reduction for the self-developed MiMo-V2.5 series, Luo Fuli, the head of Xiaomi's big 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 a break even point. The cost reduction mainly comes from hybrid attention architecture and hierarchical KV cache optimization. In response to the design goal of reducing cache hit costs by 99%, Xiaomi's inference framework has implemented hierarchical KV cache 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%. By combining the Cache Read Overlap technique between global attention modules, the system further reduces the actual overhead of cache hits. The reason for reducing the cost of basic input and output by 60% to 80% is attributed to the 1:7 inter layer sparsity ratio introduced by the model, that is, the layer ratio of global attention (GA) to sliding window attention (SWA) is 1:7. In the long text pre fill stage, the 60 layer SWA only calculates local sliding windows, which makes the overall attention computation of the MiMo-V2.5-Pro model with 70 layers equivalent to a traditional global GQA model with 10 layers. The ultra-low computing load has reduced the original inference cost, and before the price adjustment, it had reserved 2 to 3 times the profit margin for Xiaomi. Therefore, price reduction is a manifestation of structural cost reduction, rather than loss making competition. Luo Fuli said that low-cost inference services are conducive to stimulating the demand for terminal intelligence. Large model enterprises should avoid blind price wars and control actual operating expenses below the breakeven line through the underlying collaborative design of algorithms and inference systems. [Original link]