Comment by zackangelo

We're serving it around 150-200tok/s (uses our new speculative decoding implementation on a DFlash2 draft model).

https://mixlayer.com, LAUNCH-Q38-27B gets you $5 in credits if you want to kick the tires.