Qwen3-30B-A3B-4bit

quality budget ≤ +0.5% perplexity · research dataset

+63.8%

throughput, same memory

30.91

tok/s at budget 40 (5.1 GB)

cache @ 0.25σ

selected policy · quality +0.15%

savings →

Policy tournament

Every configuration measured on the real paging runtime with physical (cache-bypassed) expert reads. Residency-bias policy per arXiv:2412.00099.

budget	resident	policy	bias	tok/s	p95	reads/tok	quality
16	2.04 GB	warm	0.25σ	9.89	155.9 ms	318.4 MB	+0.32%
16	2.04 GB	none	—	9.26	160.2 ms	370.7 MB	0%
24	3.06 GB	cache	0.25σ	16.43	111.9 ms	161.4 MB	+0.15%
24	3.06 GB	cache	0.5σ	15.02	104.7 ms	126.2 MB	+1.03%
24	3.06 GB	warm	0.25σ	13.61	109.1 ms	62698.3 MB	+0.32%
24	3.06 GB	warm	0.25σ	12.9	130.1 ms	191.3 MB	+0.32%
24	3.06 GB	none	—	11.08	140.7 ms	274.8 MB	0%
24	3.06 GB	none	—	10.41	150.2 ms	278.8 MB	0%
32	4.08 GB	hybrid	0.5σ	27.01	62 ms	57 MB	+1.68%
32	4.08 GB	cache	0.5σ	24.43	64.2 ms	66.2 MB	+1.03%
32	4.08 GB	hybrid	0.25σ	19.37	84.1 ms	101.6 MB	+0.32%
32	4.08 GB	warm	0.25σ	18.1	87.5 ms	111.2 MB	+0.32%
32	4.08 GB	warm	0.25σ	17.55	91.6 ms	109 MB	+0.32%
32	4.08 GB	none	—	15.7	107.6 ms	167.2 MB	0%
32	4.08 GB	cache	0.25σ	15.62	96.4 ms	123.5 MB	+0.15%
32	4.08 GB	none	—	15.24	114.3 ms	167.2 MB	0%
32	4.08 GB	none	—	13.73	104.5 ms	67898.2 MB	0%
40	5.1 GB	cache	0.5σ	34.43	38.8 ms	22.6 MB	+1.03%
40	5.1 GB	cache	0.25σ	30.91	47.6 ms	41.8 MB	+0.15%	winner
40	5.1 GB	hybrid	0.5σ	30.35	47.8 ms	38.3 MB	+1.68%
40	5.1 GB	warm	0.25σ	25.64	60 ms	57.9 MB	+0.32%
40	5.1 GB	hybrid	0.25σ	25.4	61.2 ms	61.5 MB	+0.32%
40	5.1 GB	warm	0.25σ	25.1	59.7 ms	57.9 MB	+0.32%
40	5.1 GB	none	—	19.4	87.9 ms	125.7 MB	0%
40	5.1 GB	none	—	18.9	89.9 ms	125.7 MB	0%
40	5.1 GB	none	—	18.87	88.2 ms	118.5 MB	0%
48	6.12 GB	none	—	34.25	49.7 ms	39 MB	0%
48	6.12 GB	warm	0.25σ	29.52	47.9 ms	28.9 MB	+0.32%

Provenance: model mlx-community/Qwen3-30B-A3B-4bit · quality budget +0.5% · single-node MLX paging runtime. Rows flagged degenerate produced looping output and are excluded from selection.