Summary

SolarWM is an open framework for training video world models that support long-horizon, camera-controlled interaction. It combines a large video-data pipeline with training and inference recipes, helping researchers build environments that remain responsive and visually coherent beyond the duration of their training clips.

Key Features

The project describes 1.43 million clips totaling approximately 25TB.
The open processing pipeline provides scale-consistent camera annotations.
The framework covers Wan-5B, Wan-14B, LTX-2.5, and MiniMax-H3 variants.
The research trains on five-second clips and demonstrates substantially longer generation.
Causal student models target responsive, long-horizon world-model inference.
The release includes filtering, annotation, and training-data preparation components.
The repository announces SolarWM-H3 training, inference, weights, and pre-encoded data.
Availability varies; some backbone-specific distillation stages remain forthcoming.

About SolarWM


The project releases a corpus described as 1.43 million clips and 25TB, with camera annotations and filtering tools. A common framework supports Wan, LTX, and MiniMax-H3 backbones. Causal and few-step student models extend short-clip training toward streaming inference over much longer sessions.


SolarWM is useful for world-model research where data preparation and reproducible training are as important as the final checkpoint. The repository includes model-specific availability tables, with some distillation stages still forthcoming. Researchers should select a released backbone and recipe rather than assuming identical support across all four families.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!