Xiaomi's MiMo AI Model Slashes Compute Costs with 1:7 Sparse Attention

After Xiaomi permanently cut API prices for its MiMo-V2.5 model series, the company's head of large models, Luo Fuli, took to X to explain the algorithmic cost-cutting behind the move.

Luo revealed that even with API prices aligned to DeepSeek, Xiaomi's high-load inference engine remains profitable. The savings come from a hybrid attention architecture and hierarchical KV cache optimization.

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

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 sliding window attention (SWA) layers. During long-context prefill, the 60 SWA layers only compute local sliding windows, bringing the total attention computation of the 70-layer MiMo-V2.5-Pro model down to that of a traditional 10-layer global GQA model. This ultra-low compute load cuts raw inference costs, leaving Xiaomi with a 2x to 3x profit margin before the price cut. So the price reduction reflects structural efficiency, not a loss-leading tactic.

Luo said that low-cost inference services help stimulate demand for on-device intelligence. She advised large model companies to avoid blind price wars and instead optimize algorithms and inference systems together to keep actual operating costs below the break-even line.