Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
SolarWM is an open framework from a multi-institution research team (CUHK-SZ, SLAI, NUS, CUHK, HKUST, HKUST-GZ, NVIDIA, UCLA and MSRA) for building interactive, long-horizon video world models, covering data curation, scalable training and inference under an Apache-2.0 codebase. It converts 1.43 million canonical video clips from 14 source datasets into a unified, frame-aligned data contract covering observations, metric camera geometry, captions, quality metadata and provenance. One shared codebase trains a family of four 5B-33B backbone-native models built on Wan2.2, LTX-2.5 and MiniMax-H3, preserving each backbone's own representation. Every backbone follows the same three-stage recipe: a Stage0.5 bidirectional flow-matching stage learning video/text/camera-conditioned representations; a Stage1 stage combining teacher forcing with an AnyFlow loss so the model learns denoising and finite-step flow maps together, avoiding a separate ODE or consistency-distillation warm-up; and a Stage2 distillation stage via self-gradient forcing, training a causal student against a frozen teacher and critic on its own rollouts. Trained only on 5-second clips, the resulting causal models reportedly support real-time interaction with rollouts spanning minutes to hours, without long-sequence fine-tuning or attention-sink mechanisms. As of the September 3, 2026 release, code, the data pipeline, and full Stage0.5/1/2 weights are available for Wan2.2-5B, while Wan2.2-14B, LTX-2.5 and MiniMax-H3 ship only Stage0.5 bidirectional weights so far, with the rest "coming soon". The public SolarWM-Data release on Hugging Face and ModelScope ships license/recipe/test indexes, small examples, an annotation-only package with camera trajectories and captions, but not the full raw-WDS or latent-WDS payloads; those need separately published preencoded latents (still uploading), rebuilding raw shards from source via the annotation package, or an approved dataset access request. Code is Apache-2.0, but weights carry their own terms: LTX-2.5 derivatives use the LTX-2.x Community License and MiniMax H3 weights use a territory-restricted Community License, neither relicensed under SolarWM's terms.