Decensoring efficiency · family board
Qwen3.6-35B-A3B
6 variants · 3 methods × 2 precisions · anchor base @ BF16 · fixtures v0 · temp 0
exotic sits furthest out on the frontier — and loses only 2.1 pts of refusal reduction to NVFP4 quantization, versus 6.4 pts for heretic. The method that survives the quant that ships is the one that matters.
Efficiency frontierrefusal reduction → · drift cost ↑
basehereticexoticBF16NVFP4quant tax
BF16
NVFP4
basestock
BF16 · anchor
47.0 ref/100
tool 98.0% · KL 0.000
NVFP4
46.0 ref/100
tool 96.5% · KL 0.028
hereticTPE ablation
BF16
−81% refusals
tool 95.0% · KL 0.045
NVFP4
−74% refusals
tool 92.0% · KL 0.071
exoticours
frontier best
BF16
−87% refusals
tool 97.5% · KL 0.038
▲ quant-robust
NVFP4
−85% refusals
tool 96.0% · KL 0.052
outlined cell → full run scorecard
The NVFP4 tax
BF16 → NVFP4, per method — the quant × method interaction nobody else measuresbase
refusal Δ−1.0 /100
tool-use Δ−1.5 pts
added drift+0.028 KL
heretic
reduction Δ−6.4 pts
tool-use Δ−3.0 pts
added drift+0.026 KL
exotic
reduction Δ−2.1 pts
tool-use Δ−1.5 pts
added drift+0.014 KL