I would probably prioritize the larger models, 20t/sec seems fine, especially for what I’d be more likely to use it for (Whisper, Florence, Hunyuan, Stable Diffusion) where the LLM is mostly a go between and/or “organizer.”
But I’d probably also change how I use most of my Ai tools if I could run the largest models, so it might suddenly become something I would notice because I changed how I was using things.

