Video world models that train on seconds and simulate for hours

What it does SolarWM is a fully open foundation for building interactive video world models. It ingests 1.43 million clips from 14 datasets into a unified, frame-aligned data contract—complete with camera geometry, captions, and quality metadata—then feeds them into a three-stage training pipeline. The result is a family of causal models, spanning 5B to 33B parameters across Wan2.2, LTX-2.5, and MiniMax-H3 backbones, that can be controlled by text and camera trajectories.

The interesting bit The models are trained only on 5-second clips, yet they autoregressively roll out for minutes to hours in real time without long-sequence fine-tuning or attention-sink hacks. The trick is a staged recipe—bidirectional flow matching, teacher-forced autoregressive initialization with an AnyFlow loss, and finally distribution-matching distillation via self-gradient forcing—that works across heterogeneous backbones without requiring specialized ODE or consistency-distillation setups.

Key highlights

• Unified data contract covering 1.43 million clips from 14 datasets, decoupling source processing from training-mixture design.
• Backbone-native support for Wan2.2, LTX-2.5, and MiniMax-H3, preserving each model’s original representation and objective rather than forcing a common architecture.
• Three-stage training recipe (bidirectional flow matching → teacher-forced AnyFlow → DMD via self-gradient forcing) that avoids separate ODE or consistency-distillation pre-stages.
• Real-time, long-horizon interaction from short-clip training: causal rollouts span minutes to hours after learning only on 5-second sequences.
• Apache 2.0 licensed weights, code, and data annotations available on Hugging Face and ModelScope; full raw-WDS access requires an application form.

Caveats

• Only the Wan2.2-5B backbone has completed all three training stages released; SolarWM-14B, LTX-2.5, and MiniMax-H3 currently offer only Stage0.5 bidirectional weights, with Stage1 and Stage2 marked “Coming soon”.
• The public dataset release does not include the full raw or latent payloads by default; you must either use preencoded latent downloads, rebuild from annotation-only packages, or submit a dataset access form for the prepared raw corpus.
• Wan, LTX, and MiniMax-H3 require separate runtime environments, so mixing backbones in one environment is not supported.

Verdict Researchers and engineers experimenting with interactive video generation or world-model training should grab this, especially if they want a reproducible alternative to closed pipelines. If you need a polished, single-click consumer tool, the multi-environment setup and staged release schedule will test your patience.