SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models

TL;DR - SolarWM is an open foundation for training interactive, long-horizon video world models across heterogeneous datasets and model backbones. It matters because it standardizes data, adaptation, training, and inference while releasing the full pipeline, weights, and recipes for reproducible research.

• Unifies 1.43 million clips from 10 datasets under a frame-aligned schema containing observations, camera geometry, captions, quality metadata, selection records, and provenance.
• Adapts four models ranging from 5B to 33B parameters, based on Wan2.2, LTX-2.5, and MiniMax-H3, while retaining each backbone’s native representations and objectives.
• Uses a three-stage recipe: bidirectional adaptation, teacher-forced autoregressive initialization, and distribution-matching distillation.
• Produces causal models supporting real-time, minutes-to-hours interactive rollouts despite training on sequences only five seconds long.