Calibration
Video-MME (w/o subtitle) → CountBench
CountBench is estimated from Video-MME (w/o subtitle) with a offset logistic curve fitted on 6 models measured on both: y = 0.9177 + (0.9537 − 0.9177) / (1 + exp(−193.49·(x − 0.8259))), R² = 0.36, cross-validated error 4.3 pp. It is used for 0 estimates.
Fitted on
| Model | Video-MME (w/o subtitle) | CountBench |
|---|---|---|
| Claude Opus 4.5 | 81.4% | 90.6% |
| Gemini 3 Pro | 87.7% | 97.3% |
| GPT-5.2 | 85.8% | 91.9% |
| Kimi K2.5 | 83.2% | 94.1% |
| LFM3-VL-2B | 62.8% | 93.0% |
| Qwen3.5 397B | 83.7% | 97.2% |