CISA Dense - LongBench v2 (256–512K): leaderboard

CISA v1 AppendixA, original Nemotron3Ultra550B-A55B dense attention on LongBench v2, dense-reference input-length cohort 256–512K; fixed cohorts across precision arms, four repeats per task. BF16 and NVFP4 weights remain distinct model configurations. Zero-prompt GraphWalks cohorts remain missing, never zero-valued scores. Reported metric: Native percentage score.

Metric: Native percentage score. Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked.

Top models

#ModelScoreOverall rank
1Nemotron 3 Ultra63.5#245

No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.

Interactive version: theaggregate.ai/benchmark?slug=cisa-dense-longbench-v2-256-512k · How It Works · Data refreshed daily, snapshot 2026-11-05.