Roboflow Vision Evals — Detection mAP@50:95 (September)
Multimodal
Mean average precision across IoU thresholds 0.50 to 0.95, under the September observed protocol. One or three runs per effort tier; atomic sample count unreported.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 61.9 |
| Gemini 3.8 Flash | 59.0 |
| Gemini 3.7 Flash | 58.9 |
| Claude Opus 5.5 | 58.2 |
| Gemini 3.6 Flash | 55.3 |
| GPT-6 Sol | 54.8 |
| Gemini 3.5 Flash | 54.4 |
| Gemini 3.1 Pro Preview | 52.4 |
| Qwen3.8 Flash | 51.9 |
| GPT-6 Luna | 45.8 |
| GPT-5.6 Sol | 45.2 |
| Qwen3.7-Plus | 45.1 |
| Qwen3.5-27B | 45.1 |
| Gemini 3.5 Flash-Lite | 44.5 |
| GPT-5.6 Luna | 42.6 |
| Kimi K3 | 40.4 |
| Qwen3-VL-235B-A22B-Instruct | 40.3 |
| Claude Fable 5.1 | 39.6 |
| GPT-5.6 Terra | 39.6 |
| Muse Spark 1.2 | 39.5 |
| Muse Spark 1.1 | 39.4 |
| GLM 5V Turbo | 39.2 |
| Muse Spark 1.3 | 35.8 |
| MiMo-V2.6-Pro | 32.3 |
| Claude Fable 5 | 31.4 |
| MiMo-V2.6-Flash | 31.2 |
| Qwen3.7 Flash | 31.1 |
| Claude Opus 5 | 30.5 |
| Gemini 3 Flash Preview | 27.9 |
| Kimi K2.6 | 26.8 |
| Gemini 2.5 Pro | 24.7 |
| Grok 4.7 | 24.4 |
| GPT-5.5 | 24.0 |
| Muse Glimmer 30B | 23.0 |
| GLM-5.3 Flash | 19.0 |
| Sonnet 5 | 18.1 |
| Opus 4.8 | 16.8 |
| Grok 4.6 | 9.5 |
| Grok 4.5 | 7.9 |
| GPT-5.4 Mini | 5.9 |