Curious what's the typical switching frequency in your experiments and experience. How do you control the tradeoff of cache matching and model efficiency?