Reddit rant: 0.02TPS isn't 'running' a 5T model, you're just boiling potatoes

eightone-81 · reddit · 2026-11-19

An r/LocalLLaMA-style rant calling out low-value posts: 'I did X with a 20B/70B/120B' (old models, who cares), 'I trained a 35B' (you fine-tuned Qwen and made it worse), 'I ran 27B on potato hardware' (a brain-dead 1-bit quant on an existing engine), and 'I ran a 5T model on a potato' — at 0.02TPS with timeouts between tokens, nothing is running. You're boiling potatoes.

Original post →

More from Fun

Fun channel →