# LongCat-Video

> A 13.6B parameter open-source foundational video generation model by Meituan supporting Text-to-Video, Image-to-Video, Video-Continuation, and audio-driven avatar animation.

LongCat-Video is a foundational video generation model developed by the Meituan LongCat Team, released under the MIT License. With 13.6 billion parameters, it unifies multiple video generation tasks in a single dense-architecture model and is available on GitHub and Hugging Face for self-hosted inference.

## What It Is

LongCat-Video is an open-source video generation framework that handles Text-to-Video, Image-to-Video, Video-Continuation, and audio-driven character animation (Avatar) within a single model family. It is positioned as Meituan's first step toward world models, with a technical report published on arXiv (2510.22200). The project also includes LongCat-Video-Avatar and LongCat-Video-Avatar-1.5, specialized extensions for expressive, audio-driven human video generation.

## Architecture and Technical Design

The base model uses a dense transformer architecture (DiT) with 13.6B total and activated parameters — contrasting with mixture-of-experts alternatives that activate only a fraction of their larger parameter counts. Key architectural choices include:

- **Coarse-to-fine generation** along both temporal and spatial axes for efficient 720p, 30fps output
- **Block Sparse Attention** to improve efficiency at high resolutions
- **Multi-reward GRPO (Group Relative Policy Optimization)** for RLHF-based quality alignment
- **FlashAttention-2** enabled by default, with optional FlashAttention-3 or xformers support
- **Context parallelism** for multi-GPU inference via `torchrun`

The Avatar extensions add audio-driven lip synchronization: v1.0 uses Wav2Vec2, while v1.5 upgrades to Whisper-Large-v3 for improved accuracy, adds step distillation for 8-step fast inference, INT8 quantization for reduced VRAM, and extends to stylized domains including anime and animals.

## Supported Tasks and Workflow

Users clone the repository, set up a conda environment with PyTorch and FlashAttention, and download model weights from Hugging Face or ModelScope. Inference scripts cover:

- `run_demo_text_to_video.py` — text prompt to video
- `run_demo_image_to_video.py` — image + text to video
- `run_demo_video_continuation.py` — extend existing video clips
- `run_demo_long_video.py` — minutes-long video generation
- `run_demo_interactive_video.py` — interactive generation
- `run_demo_avatar_single_audio_to_video.py` / `run_demo_avatar_multi_audio_to_video.py` — single- and multi-character audio-driven animation
- `run_streamlit.py` — a Streamlit web UI for local use

Multi-GPU inference is supported via context parallelism (`--context_parallel_size`).

## Update: LongCat-Video-Avatar 1.5 (May 2026)

The most recent release, LongCat-Video-Avatar 1.5 (May 21, 2026), is an upgraded open-source framework for audio-driven human video generation. Notable changes from v1.0 include:

- Whisper-Large-v3 replaces Wav2Vec2 for more accurate lip synchronization
- Production-ready physical rationality and temporal stability for long videos
- Generalization to stylized domains (anime, animals, complex real-world conditions)
- Support for both single-stream and multi-stream audio inputs
- 8-step distillation inference for faster generation
- INT8 quantized DiT model option for reduced VRAM

The base LongCat-Video model was first released October 25, 2025, with LongCat-Video-Avatar following December 16, 2025. Weights are available on Hugging Face and ModelScope.

## Evaluation and Benchmarks

According to the project's internal MOS (Mean Opinion Score) benchmarks published in the README, LongCat-Video's Text-to-Video overall quality (3.38) is reported as comparable to Wan 2.2-T2V-A14B (3.35) and PixVerse-V5 (3.36), while being an open-source dense model at 13.6B parameters. For Image-to-Video, the model achieves a visual quality score of 3.27, slightly above Wan 2.2-I2V-A14B (3.23) on the same internal benchmark. These figures are vendor-published and based on Meituan's internal evaluation methodology.

## Features
- Text-to-Video generation
- Image-to-Video generation
- Video-Continuation generation
- Long video generation (minutes-long)
- Audio-driven avatar animation (single and multi-character)
- Unified multi-task architecture in a single model
- Coarse-to-fine temporal and spatial generation strategy
- Block Sparse Attention for high-resolution efficiency
- Multi-reward GRPO RLHF alignment
- FlashAttention-2/3 and xformers support
- Multi-GPU inference via context parallelism
- INT8 quantization for reduced VRAM (Avatar 1.5)
- Step distillation for fast 8-step inference (Avatar 1.5)
- Whisper-Large-v3 audio encoder for lip sync (Avatar 1.5)
- Streamlit web UI for local inference
- 720p 30fps video output

## Integrations
Hugging Face, ModelScope, FlashAttention-2, FlashAttention-3, xformers, PyTorch, Diffusers, Streamlit, Whisper-Large-v3, Wav2Vec2, ffmpeg, librosa, CacheDiT

## Platforms
CLI, API

## Pricing
Open Source

## Version
Avatar 1.5

## Links
- Website: https://github.com/meituan-longcat/LongCat-Video
- Documentation: https://arxiv.org/abs/2510.22200
- Repository: https://github.com/meituan-longcat/LongCat-Video
- EveryDev.ai: https://www.everydev.ai/tools/longcat-video