Hugging Face has released SolarWM, a fully open foundation for building interactive video world models, addressing challenges in training across heterogeneous data sources and video generator backbones. The framework spans from data preparation to long-horizon inference, aiming to standardize and reproduce results in this emerging field.
SolarWM introduces a reconfigurable multi-source data engine that converts 1.43 million canonical clips from 10 datasets into a unified, frame-aligned contract. This contract includes visual observations, metric camera geometry, captions, quality metadata, selection decisions, and provenance, while decoupling source processing from mixture construction.
The framework supports four models ranging from 5B to 33B parameters, based on Wan2.2, LTX-2.5, and MiniMax-H3, preserving each backbone's native representations and objectives. A unified three-stage training recipe combines bidirectional adaptation, teacher-forced autoregressive initialization, and distribution matching distillation.
Remarkably, the resulting causal models enable real-time interaction over rollouts lasting from minutes to hours, despite being trained only on 5-second sequences, without requiring long-sequence fine-tuning or attention-sink mechanisms. SolarWM releases its data, pipeline, recipes, weights, and framework to foster reproducible research.