A 99% cache hit cost reduction sounds exaggerated, but the SWA hierarchical KV cache can indeed increase capacity, and being below the break-even point is truly a moat.

CoinWorld reports that Luo Fuli announced on the X platform the cost reduction mechanism after the permanent price reduction of the MiMo-v2.5 API: after aligning with DeepSeek, high-load inference still maintains break-even profitability, with costs coming from hybrid attention and hierarchical KV caching. To achieve the goal of reducing cache hit costs by 99%, the Mi inference framework has optimized hierarchical KV caching for SWA, with tests showing a fivefold increase in cache capacity and an 80% cost reduction. She stated that low-cost inference can stimulate terminal demand, and enterprises should avoid malicious price cuts by coordinating algorithms and the underlying inference system to keep actual operational expenses below the break-even point.