I'm working on something from a crappy laptop, those numbers from jev can totally be matched:

    Local Latency: 0.1813 seconds

I can also run a 0.6b model on my phone faster than openai can run astra, it doesn't mean my model is useful.

and frankly for many of the kind of thing people probably want to use this for... you would want to run locally anyways.

why even bother with a network hop? build a specialized engine which does the prefill->measure cycle on local GPU/TPU/NPU with a model fine tuned for your application (e.g. gaming NPCs, autonomous driving, agricultural intelligence, drone.. target... selection, whatever)

the nice thing is that if you're skipping decode you're not as memory bandwidth bound.