SolarWM Ships Open Data and Weights for Video World Models

SolarWM landed on arXiv as a fully open release: data pipeline, training recipes, model weights, and framework for interactive video world models, pulled together into what the authors call a "reproducible and extensible foundation for interactive world-model research."

The paper frames the problem as coupling: heterogeneous datasets and different video backbones make results hard to reproduce. SolarWM's data engine "converts 1.43 million canonical clips from 10 datasets into a unified, frame-aligned contract covering visual observations, metric camera geometry, captions, quality metadata, selection decisions, and provenance," the abstract says. The project page puts the raw scale at 25 TB.

Four models ship, spanning 5B to 33B parameters across three backbones: SolarWM-Wan-5B, SolarWM-Wan-14B, SolarWM-LTX-2.5, and SolarWM-Minimax-H3. A three-stage recipe combines bidirectional adaptation, teacher-forced autoregressive initialization, and distribution matching distillation. The headline capability, per the authors: the causal models "enable real-time interaction over rollouts ranging from minutes to hours after being trained on only 5s sequences."

The GitHub repo is live, with the datasets mirrored on HuggingFace and ModelScope. It lands into a busy stretch of video-world-model work we've been tracking, including H3-World, which turns the 33B MiniMax-H3 video model into a language-controlled interactive world model. That is one of the same backbones SolarWM now packages alongside three others.