CUHK-SZ and NVIDIA, NUS, MSRA Fully Open-Source SolarWM Video World Model

A team spanning CUHK-SZ, NVIDIA, NUS and MSRA released SolarWM, an end-to-end research pipeline for interactive video world models, and open-sourced everything: paper, code, weights and data. Trained on only 5-second clips, the model supports real-time interactive rollouts lasting from minutes to hours. Scobleizer called it 'an accelerant for robotics.'

Confirmed

• SolarWM provides an open framework with a unified training recipe for building interactive video world models across diverse data sources and generator backbones
• The training corpus unifies 1.43M video clips from 14 datasets across 4 backbone architectures, roughly 25TB, all released openly
• Despite training on 5-second clips, the model sustains real-time interactive rollouts of minutes to hours while preserving responsiveness and long-horizon generation
• kwangmooyi described it as 'yet another controllable video model' while crediting the full open-sourcing

Why it matters

• Long-horizon consistency and real-time interaction are recognized challenges for video world models; SolarWM demonstrates extrapolation from very short training clips to hour-scale rollouts
• Fully releasing data, weights, source code and paper lowers the barrier to reproduction and research, and could accelerate world-model adoption in robotics and related fields

2026-09-03 ~ 2026-09-04 · 5 related posts

• [source] SolarWM: open data and scalable training for long-horizon video world models — CUHKSZ · 2026-09-03
• SolarWM goes fully open-source: 1.43M clips, 25TB pipeline for long-horizon video world models — Scobleizer · 2026-09-03
• SolarWM open-sources long-horizon video world model with dataset and weights — kwangmoo_yi · 2026-09-04
• SolarWM: Fully Open Framework Trains Long-Horizon Video World Models on 1.43M Clips — _akhaliq · 2026-09-04

1 near-duplicate retellings: kwangmoo_yi