Denza Z9S debuts CarPlay floating window mode: Apple finally figures out how to 'float' in cars

Last week we were debugging an in-car voice assistant. Every wake-up required reloading the context, pushing cold-start latency to over 2 seconds. Users straight-up called it "a senior citizen's phone." We tried pre-warming the cache, which dropped latency to 400ms, but at the cost of doubling power consumption—triggering battery life alarms again. I started wondering: is there a way to avoid full recomputation while still reusing previous results with high probability? Just saw DeepSeek-V4-Flash hitting a 98% cache hit rate, and my first thought was—if this could fit into a car system, the voice interaction experience would take off immediately.