World-model demos are everywhere, but reproducing one is painful: datasets differ in temporal scale, camera geometry and captioning style, while training recipes stay locked to a single backbone. On September 3, a team led by CUHK-Shenzhen together with NVIDIA, MSRA and six other institutions open-sourced SolarWM, an attempt to straighten all of this out in one release.
The data layer: 1.43M clips under one contract
SolarWM starts with data, not models. It converts 1.43 million canonical clips from 14 datasets into a unified, frame-aligned contract covering visual observations, metric camera geometry, captions, quality metadata, selection decisions and provenance. Source processing is decoupled from training-mixture design, so changing the mixture does not require rerunning the pipeline. The SolarWM-Data release is live on Hugging Face and ModelScope.
One recipe, three backbones
Under shared camera-conditioning, training and inference interfaces, the framework instantiates four 5B-33B models built on Wan2.2, LTX-2.5 and MiniMax-H3, each preserving its native representation and objective. The recipe has three stages: Stage0.5 learns full-clip bidirectional flow matching to establish base video, text and camera-conditioned representations. Stage1 combines teacher forcing with the AnyFlow loss, conditioning noisy target chunks on clean history while learning both denoising and finite-step flow maps, which the paper says removes the need for a separate ODE or consistency-distillation initialization before distillation. Stage2 runs distribution matching distillation via self-gradient forcing, training the causal student on its own rollout with a frozen teacher and a trainable critic.
5-second clips, hour-long interaction
The headline result: training used only 5-second sequences, yet the resulting causal models support real-time interaction with rollouts spanning minutes to hours, without long-sequence fine-tuning or attention-sink mechanisms. Code, data pipeline, recipes and weights are all out, with the code under Apache 2.0 and 249 GitHub stars on day one.
Some cold water
Three caveats in the README. First, only Wan2.2-5B has completed all three stages; the other three backbones currently ship Stage0.5 bidirectional weights only, with Stage1 and Stage2 marked Coming soon. Second, the public data release omits full raw video payloads: latent versions download per recipe, while raw-WDS must be rebuilt from the annotation package or requested via an access form. Third, openness has fine print: the code is Apache 2.0, but LTX derivatives follow the LTX community license and the MiniMax H3 license carries territory restrictions.
So what
From ForgeWM and EchoWM to SolarWM, the world-model lane has produced three full-stack open releases in two months. The competition is shifting from whose video looks better to whose data contracts and training recipes are reproducible. Whether the community picks up this cross-backbone recipe depends on when those three Coming-soon cells get filled.
References: arXiv:2609.02886; github.com/Junchao-cs/SolarWM