Calibration
SimpleVQA → VideoMMMU
VideoMMMU is estimated from SimpleVQA with a Hill curve fitted on 11 models measured on both: y = 0.0000 + (0.8786 − 0.0000)·x^6.00 / (0.34628^6.00 + x^6.00), R² = 0.98, cross-validated error 10.3 pp. It is used for 2 estimates.
| Estimated model | SimpleVQA | VideoMMMU | Source |
|---|---|---|---|
| Step 3.7 Flash | 79.2% | 87.2% | estimated ± 10.3 pp, low confidence |
| Muse Spark Open | 73.6% | 86.9% | estimated ± 10.3 pp, low confidence |
Fitted on
| Model | SimpleVQA | VideoMMMU |
|---|---|---|
| Claude Opus 4.5 | 65.7% | 84.4% |
| dots3-note Preview | 72.5% | 86.8% |
| Gemini 3 Pro | 73.2% | 87.6% |
| GPT-5.2 | 55.8% | 85.9% |
| Kimi K2.5 | 71.2% | 86.6% |
| LFM3-VL-2B | 36.3% | 49.6% |
| Qwen3.5 397B | 67.1% | 84.7% |
| Qwen3.6-27B | 56.1% | 84.4% |
| Qwen3.6-35B-A3B | 58.9% | 83.7% |
| Qwen3.7 Plus | 81.7% | 85.4% |
| Qwen3.8 Max | 75.0% | 88.7% |