Hacker News Reader

Most LLMs cannot run efficiently on current NPUs (except for prefill stage), the hardware was built for a different kind of ML workload.

Decision models are prefill only

That is most of the explanation for their speed, even.