🏆 ICML 2026 — Efficient Qwen Leaderboard

Minimizing Inference Latency for Qwen3.5-4B on A10G

Hardware: NVIDIA A10G (ml.g5.xlarge) | Ranked by average speedup over baseline

Baseline (unoptimized Qwen3.5-4B): Short: 2,582 ms | Medium: 5,441 ms | Long: 6,576 ms | Avg: 4,866 ms

#	Team	Avg Speedup	Short (64t)	Medium (2048t)	Long (8192t)	MMLU-Pro ≥62.1%	IFEval ≥81.4%	GPQA-Diamond ≥63.0%
1	AFM-as4vvw34	5.576x	333 ms	835 ms	2,678 ms	✅ 0.628	✅ 0.833	✅ 0.636
2	AFM-69568ssx	3.404x	620 ms	1,312 ms	3,456 ms	✅ 0.648	✅ 0.833	✅ 0.667
3	AFM-hnknxz5w	3.401x	622 ms	1,317 ms	3,424 ms	✅ 0.658	✅ 0.845	✅ 0.707
4	AFM-xpvr9w7k	3.013x	668 ms	1,594 ms	3,729 ms	✅ 0.646	✅ 0.821	✅ 0.636
5	AFM-pysuua5t	2.799x	774 ms	1,825 ms	3,161 ms	✅ 0.652	✅ 0.857	✅ 0.636
6	AFM-wp5spjm8	2.432x	1,154 ms	2,308 ms	2,435 ms	✅ 0.649	✅ 0.857	✅ 0.687
7	AFM-ayxvmqbj	2.103x	1,124 ms	2,500 ms	3,580 ms	✅ 0.663	✅ 0.821	✅ 0.657
8	AFM-z9qv547h	2.103x	1,125 ms	2,500 ms	3,582 ms	✅ 0.644	✅ 0.821	✅ 0.677
9	AFM-gv7e2ebx	2.103x	1,124 ms	2,499 ms	3,579 ms	✅ 0.647	✅ 0.821	✅ 0.636
10	AFM-npq35pxt	1.570x	1,584 ms	3,384 ms	4,468 ms	✅ 0.669	✅ 0.845	✅ 0.636
11	AFM-rc3csarb	1.481x	1,743 ms	3,652 ms	4,467 ms	✅ 0.672	✅ 0.869	✅ 0.727
12	AFM-newrkm40	0.999x	2,584 ms	5,444 ms	6,577 ms	✅ 0.688	✅ 0.845	✅ 0.687
13	Baseline	0.999x	2,585 ms	5,444 ms	6,577 ms	✅ 0.685	✅ 0.857	✅ 0.697