Long-context language understanding on LongBench (16 Subtasks + Avg) benchmark leaderboard · SOTA2 Research

Long-context language understanding on LongBench (16 Subtasks + Avg)

44.492Wiki ScoreRTPurbo

Evaluation Results

Method	Links
RTPurbo2026.05		44.49	54.74	35.29	53.34	65.56	40.22	31.92	24.6	14	90.85	84.5	60	34.56	34.08	99.75	100	54.24
SnapKV2026.05		42.56	53.1	38.14	52.7	65.5	39.3	16.11	17.68	11.03	82.03	77	57.21	35	25.76	99	99.75	50.74
RTPurbo2026.05		42.56	53.1	36.46	51.75	64.97	40.43	31.76	24.41	13.88	89.1	83	55.5	34.41	33.47	99	99	53.3
Full Attn2026.05		42.08	54.64	38.3	52.89	65.99	39.25	31.89	23.77	13.55	89.93	85.5	61.5	35.08	27.61	99.75	99	53.8
RazorAttn2026.05		40.31	50.92	29.15	53.04	65.9	39.72	31.61	23.59	13.85	89.26	85	59.5	35.01	32.05	99	99.75	52.98
Minference2026.05		40.22	52.21	32.12	52.09	64.92	38.43	31.77	23.6	13.42	51.95	82.5	32.25	34.05	25.76	99	100	48.39
Quest2026.05		39.11	53.27	31.64	48.22	62.8	38.27	30.18	23.55	13.24	79.88	81	59	35.8	24.31	98.83	91.92	50.69
FlexPrefill2026.05		36.48	52.76	35.71	51.69	65.86	37.44	31.63	24	13.1	90.59	80.5	50.25	32.85	24.93	93.5	69.5	49.42