RULER: leaderboard
13 synthetic long-context tasks from NVIDIA (2024) in four groups (needle-in-a-haystack retrieval, variable tracing, word-frequency aggregation, distractor-padded QA) generated at 4K to 128K tokens.
Metric: Avg Accuracy (%). Source: github.com. Status: saturated. 45 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Jamba 1.5 Large | 96 |
| 2 | Gemini 1.5 Pro | 95.8 |
| 3 | Qwen2.5-14B-Instruct-1M | 95.7 |
| 4 | Qwen 3 235B A22B | 95 |
| 5 | Qwen 3 14B | 94.6 |
| 6 | Jamba 1.5 Mini | 93.9 |
| 7 | Qwen 3 32B | 93.7 |
| 8 | Qwen2.5-7B-Instruct-1M | 91.8 |
| 9 | Qwen 3 30B A3B | 91.6 |
| 10 | GPT-4 Preview (1106) | 91.6 |
| 11 | Qwen 3 8B | 89.1 |
| 12 | Command R | 88.9 |
| 13 | Command-R+ | 87.9 |
| 14 | Mistral Large 2 (Nov) Instruct (2411) | 86 |
| 15 | Qwen 3 4B | 85.2 |
Interactive version: theaggregate.ai/benchmark?slug=ruler · How It Works · Data refreshed daily, snapshot 2026-09-05.