Benchmarks
Published video Elo and a test protocol for interactive world models.
Published video Elo
Artificial Analysis AA-Video-I2V v1.1: image-to-video with audio, rated by blind human preference. Checked October 15, 2026.
| Exact tested model | Elo | 95% interval | Samples | Source |
|---|---|---|---|---|
| MiniMax H3 (768p) | 1141 | 1124–1158 | 4,009 | Artificial Analysis |
| MiniMax H3 Max | 1124 | 1107–1141 | 3,255 | Artificial Analysis |
Scores apply to the listed model and task. The 95% intervals overlap. Compare Elo within the same benchmark; world consistency and interaction latency need separate tests.
Interactive model test protocol
Input latency
Time from an input to the first visible response. Report p50, p95, network conditions, cold starts, and stalls.
Generation & playback
Report generation fps, delivered fps, dropped frames, and playable seconds per elapsed second. Include decoding and transport.
World consistency
Use repeated trials with matched inputs and action scripts. Blind reviewers assess geometry, identity, object permanence, and prompt adherence.
Cost
Record billed duration, warmup, idle charges, and cost per successful run, session hour, and output hour.
Metric definitions
Resolution labels specify native or upscaled output. Frame rates distinguish generated, delivered, and interpolated frames. Throughput depends on hardware and may exclude decoding.
Evaluate action-controlled worlds, prompt-directed video, and 3D generators on separate tasks.
Test template
Record model versions, inputs, hardware, timestamps, billing units, measurements, and output links in the blank template.