experiments/analysis/tables/t4_quantization.md

Qwen3-8B — target quant (q4/q5/q8) x drafter

Target quant	Drafter	n	tok/s	x vs solo	alpha	tau
q4	Vanilla-1.7B	1474	75.5	1.41x	0.725	173
q4	EAGLE-3	1474	74.4	1.39x	0.440	141
q4	DFlash-F16	1474	33.4	0.63x	0.009	7
q4	DSpark p=0.0	1474	87.6	1.64x	0.616	162
q4	DSpark p=0.2	1474	86.7	1.62x	0.621	162
q4	DSpark p=0.4	1474	85.4	1.60x	0.646	161
q4	DSpark p=0.6	1474	80.7	1.51x	0.714	154
q5	EAGLE-3	1474	68.3	1.46x	0.439	140
q5	DFlash-F16	1474	30.8	0.66x	0.009	7
q5	DSpark p=0.0	1474	80.6	1.72x	0.613	162
q8	EAGLE-3	1474	52.8	1.63x	0.432	140
q8	DSpark p=0.0	1474	62.3	1.93x	0.611	162

Gemma 4 12B — target quant x drafter

Target quant	Drafter	n	tok/s	x vs solo	alpha	tau
q4	DFlash-F16	1455	72.2	2.09x	0.657	156
q4	DFlash-Q4	1455	78.4	2.26x	0.647	155
q4	DFlash-Q8	1455	76.6	2.21x	0.656	156
q4	MTP	1455	82.7	2.39x	0.739	164
q5	DFlash-F16	1449	64.5	2.16x	0.659	157
q5	MTP	1449	75.6	2.53x	0.745	165
q8	DFlash-F16	1452	45.7	2.30x	0.658	156
q8	MTP	1452	54.0	2.71x	0.745	164

Gemma-4 Q4 — draft quantization effect (DFlash drafts, final runs)

Drafter (draft quant)	n	tok/s	x vs solo	alpha	tau	max VRAM (MiB)
DFlash-F16	1455	72.2	2.09x	0.657	156	10766
DFlash-Q4	1455	78.4	2.26x	0.647	155	9730
DFlash-Q8	1455	76.6	2.21x	0.656	156	10043

F16 = 1.47 GB draft, Q4_K_M = 0.44 GB, Q8_0 = 0.79 GB (model-hashes.json).