LongCat-Video

Minutes-long video without color drift is the differentiator for LongCat-Video, a 13.6B-parameter foundation model from Meituan's LongCat team released in late 2025. One model handles text-to-video, image-to-video, and video continuation, a unification that comes from pretraining on the continuation task itself, and output runs at 720p and 30 fps. Efficiency comes from a coarse-to-fine generation strategy along both the temporal and spatial axes plus block sparse attention at high resolutions, while quality is tuned with multi-reward GRPO, a reinforcement learning from human feedback scheme. Setup is a conda environment with Python 3.10, PyTorch built for CUDA 12.4, and Flash Attention 2, with single and multi-GPU inference paths; the follow-up LongCat-Video-Avatar-1.5 release adds INT8 quantization to cut VRAM use. Code and weights are MIT licensed, an unusually permissive choice for a video foundation model, and the repository has about 5,300 stars. Long-form content generation and world-model research are the stated targets, and commercial use is unrestricted.