13B-14B models hit 22-53 tok/s at Q4-Q6, a 35B-A3B MoE runs via expert offload, and Flux runs at FP8. Where 16GB beats 12GB, where it trails 24GB, and the best cards at this tier.