Roboflow Vision Evals
Multimodal · 2026-07-22
Roboflow Vision Evals compares vision-language models on 513 real-image samples across object detection, counting, identification, OCR, data extraction, and visual reasoning. The headline score is the unweighted mean of the six task-level headline scores.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash | 87.1 |
| Gemini 3.1 Pro Preview | 86.3 |
| Gemini 3.6 Flash | 82.8 |
| Claude Fable 5 | 82.2 |
| Muse Spark 1.1 | 79.1 |
| GPT-5.6 Sol | 78.3 |
| Gemini 3 Flash Preview | 77.1 |
| GPT-5.6 Luna | 76.8 |
| GPT-5.5 | 76.4 |
| GPT-5.6 Terra | 76.2 |
| Gemini 3.5 Flash-Lite | 70.3 |
| Gemini 2.5 Pro | 69.1 |
| Sonnet 5 | 68.7 |
| Opus 4.8 | 68.1 |
| Qwen3-VL-235B-A22B-Instruct | 68.0 |
| Qwen3.7-Plus | 68.0 |
| Kimi K3 | 67.4 |
| GPT-5.4 Mini | 67.3 |
| GLM 5V Turbo | 67.2 |
| Qwen3.5-27B | 66.2 |
| Kimi K2.6 | 57.6 |