LongCat-Video

Overview

LongCat-Video is a 13.6-billion-parameter video generation model built on a Diffusion Transformer (DiT) architecture, released by Meituan's LongCat team in October 2025. It unifies three video generation tasks — text-to-video, image-to-video, and video continuation — within a single model framework, eliminating the need for separate task-specific models. The model generates 720p video at 30 frames per second within minutes, using a coarse-to-fine generation strategy along both temporal and spatial dimensions. The publisher positions LongCat-Video as 'a first step toward world models,' emphasizing long-form video generation with high temporal consistency across scenes up to 5 minutes in duration. In terms of video quality, the model applies multi-reward RLHF during training to optimize visual quality, motion naturalness, and text-video alignment. It is released with inference code and model weights under the meituan-longcat GitHub organization.

Architecture

LongCat-Video is built on a Diffusion Transformer (DiT) backbone, described in the technical report (arXiv:2510.22200), with two key efficiency innovations. First, the coarse-to-fine generation strategy first produces a low-resolution temporal outline of the video, then refines it spatially and temporally in subsequent passes — avoiding the cost of generating full-resolution frames for every timestep. Second, Block Sparse Attention limits attention computation to spatially and temporally nearby tokens, reducing the quadratic complexity of standard attention. This is particularly important for long videos (up to 5 minutes / 9,000 frames) where naive full attention would be computationally prohibitive. The model also introduces a temporal consistency mechanism spanning adjacent chunks to prevent frame-to-frame flickering in extended generations. The DiT model processes video latents in a compressed space, decoding to 720p at the final output stage.

Training

The model was trained using multi-reward RLHF (Reinforcement Learning from Human Feedback) with multiple reward models optimizing for visual quality, motion smoothness, temporal consistency, and text-video alignment. This differs from standard image-generation RLHF by incorporating temporal-video-specific evaluation dimensions that must assess both individual frame quality and cross-frame coherence. Training details including data scale and compute resources are described in the accompanying technical report (arXiv:2510.22200).

Deployment

LongCat-Video model weights and inference code are available on the meituan-longcat GitHub repository and HuggingFace. The model is also listed on ModelScope for Chinese-region access. For deployment, the DiT backbone can be served using standard diffusion model inference pipelines with appropriate GPU memory. The LongCat-Video project page provides documentation for configuration and usage across the three supported tasks (T2V, I2V, video continuation). Video generation at 720p/30fps typically requires server-grade GPU hardware with sufficient VRAM for the 13.6B-parameter model.

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

• Text-to-video generation
• Image-to-video generation
• Video continuation

Access & License

Open source; verify the repository license before commercial use.

Sources

• LongCat-Video GitHub Repository

  Primary source · Accessed 2026-07-16

  Primary repository under the meituan-longcat organization.

• LongCat-Video Project Page

  Publisher documentation · Accessed 2026-07-16

  Publisher project page.

Related comparisons

• LongCat-Video and Seedance 2.5: Open-Weight vs Closed Video Generation Compared

  A source-based comparison of LongCat-Video (13.6B open-weight DiT, 720p/30fps, text/image-to-video, up to 5-minute generations, self-hostable) and Seedance 2.5 (ByteDance's closed video flagship, 30s native output, 50 multimodal references, built-in audio). Covers architecture, duration, resolution, pricing, licensing, and deployment.

• LongCat-Video and Seedance 2.0: Open-Weight vs Closed Video Generation Compared

  A source-based comparison of LongCat-Video (13.6B open-weight DiT, 720p/30fps, up to 5-minute generations, self-hostable) and Seedance 2.0 (ByteDance's closed video model, 4-15s clips, native audio, 480p-1080p). Covers architecture, duration, audio, pricing, licensing, and deployment — the predecessor of Seedance 2.5.

• LongCat-Video and Wan 2.2: Two Open-Weight Video Generation Models Compared

  A source-based comparison of LongCat-Video (13.6B DiT, 720p/30fps, up to 5-minute generations) and Wan 2.2 (Alibaba's open-weight MoE video model, 27B total/14B active, plus 5B consumer-GPU model). Both open licenses; covers architecture, duration, resolution, deployment, and licensing.

FAQ

Where can the primary source for LongCat-Video be found?

The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Video.

Is LongCat-Video available through the LongCat API?

LongCat-Video is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.

Is this page an official LongCat or Meituan page?

No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.

How often is this page updated?

This page was last verified on 2026-07-16. Content is reviewed when new publisher documentation or model releases become available.