SolarWM — open data and training code for long-horizon video world models

What is it?

SolarWM publishes the parts of video world-model research that usually stay private: the data pipeline, the training code and the weights. A reconfigurable multi-source data engine folds 1,425,694 video clips from 14 datasets into one frame-aligned format with camera measurements and quality labels. Four models were trained on it — Wan2.2-5B, Wan2.2-14B, LTX-2.5 and MiniMax-H3 — spanning 5B to 33B parameters, each keeping its backbone's native design.

How does it work?

Training runs in three stages that turn a bidirectional video model into a causal, camera-controlled one. Stage0.5 learns full-clip bidirectional flow matching and builds the video, text and camera-conditioned representation. Stage1 mixes teacher forcing with denoising so no separate ODE initialization is needed. Stage2 distills the distribution with self-gradient forcing against a frozen teacher. The paper reports that models trained only on 5-second sequences roll out for minutes to hours without long-sequence fine-tuning or attention-sink mechanisms.

Why does it matter?

Reproducing a video world model normally means rebuilding a million-clip corpus first, which prices most labs out. SolarWM ships that corpus, its source registry and the reconstruction tools, with the code under Apache-2.0 while each video source keeps its own licence. Note the release is partial: all Wan2.2-5B stages are public, while the 14B, LTX-2.5 and MiniMax-H3 variants currently have Stage0.5 weights only.

Who is it for?

video and world-model researchers

Try it

python -m pip install -e . && solarwm environment probe

Tags

• paper
• dataset
• world-model
• video-generation
• open-data
• camera-control
• distillation
• apache-2-0
• arxiv
• ltx-2-5
• minimax-h3