it's so interesting how often a bare tok/s number blows up on here, and some of them feel like a scam. i see a huge improvement label, i load the exact recipe on my node, run it across the context, and once the window fills it lands right on my baseline.
a fresh chat number with no context is the easiest number in local ai to make look good. speed drops as the window fills and spec decoding pays the most on an empty chat, so one fresh number with no depth tells you almost nothing about the agent you'll actually run for hours.
the algo can't check any of that, it ranks engagement, so the hype gets the reach. from what i've watched, all it takes is 4 to 8 accounts that reply early and boost each other every time, and an unverified number ends up all over the timeline.
so before you believe a speed claim, ask for the context depth, the baseline on the same card and build, and whether spec decoding was on. if a post can't show you those, treat the number as marketing until you've run it yourself.