LongCat-Video
Dataset & Modelmeituan-longcat/LongCat-Video
13.6B foundational model for text-to-video, image-to-video and long video generation.
Overview
LongCat-Video is Meituan's 13.6B foundational video generation model unifying text-to-video, image-to-video and video continuation in a single architecture. It efficiently creates minutes-long 720p videos without quality drift, trained with multi-reward RLHF, and includes an Avatar variant for audio-driven character animation.
README Preview
# LongCat-Video\n\n\n \n\n\n\n\n\n \n \n \n \n\n\n \n \n \n \n \n\n\n \n \n \n \n\n \n\n\n## Model Introduction\nWe introduce LongCat-Video, a foundational video generation model with 13.6B parameters, delivering strong performance across *Text-to-Video*, *Image-to-Video*, and *Video-Continuation* generation tasks. It particularly excels in efficient and high-quality long video generation, representing our first step toward world models.\n\n### Key Features\n- 🌟 **Unified architecture for multiple tasks**: LongCat-Video unifies *Text-to-Video*, *Image-to-Video*, and *Video-Continuation* tasks within a single video generation framework. It natively supports all these tasks with a single model and consistently delivers strong performance across each individual task.\n- 🌟 **Long video generation**: LongCat-Video is natively pretrained on *Video-Continuation* tasks, enabling it to produce minutes-long videos without color drifting or quality degradation.\n- 🌟 **Efficient inference**: LongCat-Video generates $720p$, $30fps$ videos within minutes by employing a coarse-to-fine generation strategy along both the temporal and spatial axes. Block Sparse Attention further enhances efficiency, particularly at high resolutions\n- 🌟 **Strong performance with multi-reward RLHF**: Powered by multi-reward Group Relative Policy Optimization (GRPO), comprehensive evaluations on both internal and public benchmarks demonstrate that LongCat-Video achieves performance comparable to leading open-source video generation models as well as the latest commercial solutions.\n\nFor more detail, please refer to the comprehensive [***LongCat-Video Technical Report***](https://arxiv.org/abs/2510.22200).\n\n## 🎥 Teaser Video\n\n\n \n\n\n## 🔥 Latest News!!\n- May 21, 2026: 🚀 We release [***LongCat-Video-Avatar-1.5***](https://meigen-ai.github.io/LongCat-Video-Avatar-1.5-Page/), an upgraded open-source framework for audio-driven human video generation. v1.5 replaces Wav2Vec2
FAQ (1)
How-toDoes the LongCat-Video-Avatar 1.5 base model include GRPO training, and how to run non-DMD baseline inference?
Yes. The released base_model and base_model_int8 weights are merged checkpoints that already include the GRPO (RLHF) training stage. The lora/dmd_lora.safetensors file is only a DMD2 acceleration adapter trained on the *non-GRPO* base version — it provides step distillation, not the GRPO quality improvements.
To evaluate the non-DMD baseline quality of v1.5, run avatar-v1.5 without --use_distill and set inference steps to 40 with the default scheduler/config. When using --use_distill (DMD), 8 steps is enough, and per the maintainers the DMD-distilled setup gives better stability than the non-distilled baseline.