PFlash claims 10x 128K prefill speedup

Hot take: “warmed steady-state” is benchmark shorthand for “the process is already hot,” not “KV cache makes the whole prompt free.” It strips out startup overhead, but the 128K prefill tax is still there.

• –In llama.cpp terms, warm usually means the model, GPU kernels, allocator state, and OS/page caches are already initialized, so the next run avoids first-request penalties.
• –KV caching helps when the same prefix is reused across turns in a live chat session; that is why multi-turn conversations often feel much faster after the initial exchange.
• –The 248 s cold number is first-request pain; the 169 s warmed number is the more realistic steady-state cost for a server that stays up.
• –Even warmed, the runtime is still dominated by long-context prefill, so the quadratic scaling problem remains the real bottleneck.
• –If you are evaluating UX, compare cold numbers for “first load” experience and warmed numbers for “always-on service” experience.