DropKV Reference - RULER 64K (full KV cache): leaderboard
Full KV-cache rows only (no eviction; the unmodified model), lm-evaluation-harness on one A100-80GB. RULER at 64K sequence length, mean of the 11 tasks the paper reports (NIAH single 1-3, multi-key 1-2, multi-query, multi-value, variable tracking, frequent-word extraction, QA-1, QA-2). Eviction-method rows change the evaluated system and are not published. Mistral-7B-512K is the paper's short name for MegaBeam-Mistral-7B-512K. Metric: Accuracy (%). Higher is better. Source: https://arxiv.org/html/2610.22130v1
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 8B Instruct | 91.35 | #1029 |
| 2 | Qwen2.5-7B-Instruct-1M | 88.73 | #854 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=dropkv-reference-ruler-64k-full-kv-cache · How It Works · Data refreshed daily, snapshot 2026-10-18.