Prefil of local models vs opus and astra

Why does no one talk about what the prefil speeds of these API providers are vs running locally. People with sparks or strix halos only seem to focus on decode without considering how much slower it is because of slow pp. Are there any benchmarks / figures of how fast the APIs process input? subm

中文全文翻译正在补齐，可先查看下方英文原文。

Why does no one talk about what the prefil speeds of these API providers are vs running locally. People with sparks or strix halos only seem to focus on decode without considering how much slower it is because of slow pp. Are there any benchmarks / figures of how fast the APIs process input? subm