Our slowest image model is also the fastest per step. qwen-image takes 0.27 s a step and needs 50 of them. flux2-klein takes 0.29 s and needs 4. So 13.4 s an image against 1.2 s, same card, same…

lama-3.1-8b, time to first token. We land at 205 ms p95, measured end-to-end on EcoHash, one RTX Pro 6000. Peers on this chart run 520 to 1320 ms, per Artificial Analysis. Purple dot is us. Both axes…