That’s what I suspect, too, but I’m not entirely sure in my research so far. The question I am still unsure about: Is it as costly in running, or is the real costly part “just” the “training our model” part? I wondered that, because when I was messing around, things like generative text models could run on my potato PC with a bit of python scripting without too much issue, even if not ideally - as long as I had the already trained dataset downloaded.