PatternEval (Non-Thinking): leaderboard
Metric: Accuracy (%; correctness of the final user-visible answer under the source task's criteria, judged by Qwen3-Max; non-thinking inference; PatternEval: 2,415 failure-enriched image-prompt pairs in three task families (visual perception and grounding, OCR and structured images, multimodal knowledge reasoning); official default decoding with only the reasoning control changed). Source: arxiv.org. Saturation forecast: Around March 2027. 25 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.7 Max (Non-reasoning) | 67.07 |
| 2 | Seed 2.0 Pro (Non-reasoning) | 66.85 |
| 3 | Kimi K2.6 (Non-reasoning) | 66.47 |
| 4 | Qwen 3.7 Plus (Non-reasoning) | 66.28 |
| 5 | Qwen 3.5 397B A17B (Non-reasoning) | 61.96 |
| 6 | Kimi K2.5 (Non-reasoning) | 61.33 |
| 7 | Claude Opus 4.8 | 61.12 |
| 8 | GPT-5.4 (Non-reasoning) | 60.66 |
| 9 | Claude Opus 4.6 | 59.45 |
| 10 | GPT-5.5 (Non-reasoning) | 59.2 |
| 11 | Qwen 3.5 122B A10B (Non-reasoning) | 59.2 |
| 12 | Qwen 3.5 27B (Non-reasoning) | 59.07 |
| 13 | Qwen 3.6 27B (Non-reasoning) | 57.33 |
| 14 | Qwen 3.5 35B A3B (Non-reasoning) | 57.02 |
| 15 | Qwen 3.6 35B A3B (Non-reasoning) | 56.22 |
Interactive version: theaggregate.ai/benchmark?slug=patterneval-non-thinking · How It Works · Data refreshed daily, snapshot 2026-09-29.