SolarWM open-sources long-horizon video world model with 1.43M clips and 25TB dataset
kwangmoo_yi · x · 2026-09-04
A team spanning CUHK-SZ, NVIDIA, NUS and others released SolarWM, an open video world model with paper, code, weights and data all public.
- Long-horizon real-time inference: trained only on 5-second clips, it enables minute- to hour-scale real-time inference while keeping responsive interaction and coherent world evolution over extended rollouts
- Open data pipeline: 1.43M clips / 25TB with a full processing stack — scale-consistent camera annotations, fine-grained filtering, and multiple training recipes
- One framework, many backbones: supports SolarWM-Wan-5B, SolarWM-Wan-14B, SolarWM-LTX-2.5 and SolarWM-Minimax-H3
Related event: CUHK-SZ and NVIDIA, NUS, MSRA Fully Open-Source SolarWM Video World Model(5 posts)→
More from Multimodal
- FastVideo open-sources MiniMax FastH3 Omni with training code, 12x faster on Blackwell — multimodalart · 2026-10-21
- Megalith 008: stunning AI-generated colossal architecture series — azed_ai · 2026-10-21
- Higgsfield offers $25K prizes for open-source ComfyUI rebuilds of its apps — MuziqueComfyUI · 2026-10-21
- Stunning AI-generated elderly portrait blurs the line with real photography — LudovicCreator · 2026-10-21
- Writer with no drawing skills uses AI to publish third episode of original manga — OriginalMangaStudio · 2026-10-21
- FastH3 Omni open-sourced: reference-conditioned video with synced audio, 5.6-11.9x lower latency — mnmunknown · 2026-10-21