CISA Dense - LongBench v2 (Aggregate): leaderboard
CISA v1 Table1, original Nemotron3Ultra550B-A55B NVFP4 weights with dense attention, four repeats per task, top-p1 and65536-token output cap. GraphWalks uses BF16 KV/Mamba caches and FP16 SSM state. Unbound Inkling-Small deployment identity is retained for follow-up, not inferred from its short label. Reported metric: Native percentage score.
Metric: Native percentage score. Source: arxiv.org. 1 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nemotron 3 Ultra (NVFP4) | 61.38 |
Interactive version: theaggregate.ai/benchmark?slug=cisa-dense-longbench-v2-aggregate · How It Works · Data refreshed daily, snapshot 2026-11-05.