The speed of model evolution is crazy. Agentic workloads made kv caching the bottleneck, but now models are evolving to drastically reduce kv.

Chart from @deepseek_ai.