Roboflow Vision Evals - Visual Reasoning
Multimodal · 2026-07-22
The Roboflow Vision Evals visual-reasoning task measures exact-match accuracy on 23 image questions that require spatial or relational reasoning.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 91.3 |
| Claude Fable 5 | 87.0 |
| Gemini 3.5 Flash | 87.0 |
| GPT-5.6 Luna | 87.0 |
| GPT-5.5 | 82.6 |
| GPT-5.6 Terra | 82.6 |
| Gemini 3.6 Flash | 78.3 |
| Gemini 3 Flash Preview | 78.3 |
| GPT-5.4 Mini | 78.3 |
| GPT-5.6 Sol | 73.9 |
| Muse Spark 1.1 | 73.9 |
| Opus 4.8 | 60.9 |
| Gemini 2.5 Pro | 60.9 |
| Sonnet 5 | 56.5 |
| Gemini 3.5 Flash-Lite | 52.2 |
| Kimi K3 | 47.8 |
| GLM 5V Turbo | 43.5 |
| Qwen3.5-27B | 43.5 |
| Qwen3.7-Plus | 43.5 |
| Qwen3-VL-235B-A22B-Instruct | 43.5 |
| Kimi K2.6 | 30.4 |