Calibration
Video-MME (with subtitle) → CharXiv
CharXiv is estimated from Video-MME (with subtitle) with a Hill curve fitted on 10 models measured on both: y = 0.3734 + (1.2000 − 0.3734)·x^6.00 / (0.85282^6.00 + x^6.00), R² = 0.67, cross-validated error 4.9 pp. It is used for 1 estimate.
| Estimated model | Video-MME (with subtitle) | CharXiv | Source |
|---|---|---|---|
| MiniMax M3 | 85.4% | 78.8% | estimated ± 4.9 pp, high confidence |
Fitted on
| Model | Video-MME (with subtitle) | CharXiv |
|---|---|---|
| Claude Opus 4.5 | 77.6% | 68.5% |
| Gemini 3 Pro | 88.4% | 81.4% |
| GPT-5.2 | 86.0% | 82.1% |
| Kimi K2.5 | 87.4% | 77.5% |
| MiMo-V2.5 | 87.7% | 81.0% |
| Qwen3.5 397B | 87.5% | 80.8% |
| Qwen3.6-27B | 87.7% | 78.4% |
| Qwen3.6-35B-A3B | 86.6% | 78.0% |
| Qwen3.7 Plus | 88.0% | 85.9% |
| Qwen3.8 Max | 90.4% | 93.5% |