Foundational video generation · 13.6B parameters
Long video is a
continuation problem.
LongCat-Video handles text-to-video, image-to-video and video-continuation in one model. Because it is natively pretrained on continuation, the project says it produces minutes-long videos without color drifting or quality degradation.
One model · three ways in
illustrative strip- Text → videoT2V
- Image → videoI2V
- Video → more videonative pretraining task
720p · 30 fpsgenerated within minutes — the project’s figure
Why the long version stays affordable
three mechanisms the README namesTwo axes, not one
Generation runs coarse-to-fine along both the temporal and the spatial axis, which is how 720p at 30 fps stays within minutes.
Cheaper at high resolution
Block Sparse Attention further improves efficiency, with the largest effect at high resolutions.
Tuned against several rewards
Reinforcement learning from human feedback uses multi-reward Group Relative Policy Optimization rather than a single score.
“Our first step toward world models.”
No benchmark score is reproduced here; the project’s comparisons live in its technical report. The strip above is illustrative.