CISA Dense - Terminal-Bench Hard (Capability): leaderboard
CISA v1 AppendixA Table6, original dense models on Terminal-Bench Hard, temperature1. Native tasks/repeats are100/16 AA-LCR,500/3 SWE-bench,89/8 Terminal-Bench2.0,48/8 Terminal-BenchHard and220/2 GDPval. Terminal-Bench uses context summarization; BF16 weights use FP8 KV and NVFP4 weights use BF16 KV. GDPval compares against Kimi-K2.5-Thinking vadamsr2 with four Gemini3.1ProPreview judgments per rollout; missing judgments count as losses, all440 rollouts retained. Precision and missingness contracts are explicit. Reported metric: pass@1 / resolved (%).
Metric: pass@1 / resolved (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nemotron 3 Ultra (BF16, FP8 KV) | 37.24 | |
| 2 | Nemotron 3 Ultra (NVFP4) | 35.94 |
Interactive version: theaggregate.ai/benchmark?slug=cisa-dense-terminal-bench-hard-capability · How It Works · Data refreshed daily, snapshot 2026-11-05.