What is LongCat-Video?
LongCat-Video is an open-source video generation model from Meituan's LongCat team, released under the permissive MIT License. A single 13.6B-parameter model covers text-to-video, image-to-video, and video-continuation, and it is built for minute-long clips at 720p/30fps without the color drift that usually wrecks long generations.
It is a dense model (not mixture-of-experts). On Meituan's internal benchmark it scores competitively with other open models while approaching proprietary systems like Veo3 (see benchmarks below). Weights are published on Hugging Face.
Who it's for: teams and creators who want to self-host video generation — short-form clips, B-roll, or (via the avatar branch) audio-driven talking-head video — instead of paying per call for a closed API like Veo.
Can your GPU run it?
Short answer: not on a typical consumer card without quantization.
Meituan never published a minimum-VRAM number — the community has been asking since launch (issue #20, still open). What the field reports tell us:
- A user hit out-of-memory on a 48GB GPU (issue #79).
- The community fast fork (
xlite-dev/longcat-video-fast) adds 4/8-bit quantization specifically to fit GPUs "< 48GB memory."
So plan on a 48GB+ GPU for the full-precision model, or use a quantized build for 24GB-class cards. Multi-GPU inference is supported via context-parallelism, though some users report failures on 4 cards that worked on a single card (issue #27).
Install LongCat-Video
git clone --single-branch --branch main https://github.com/meituan-longcat/LongCat-Video
cd LongCat-Video
conda create -n longcat-video python=3.10 && conda activate longcat-video
pip install torch==2.6.0+cu124 torchvision==0.21.0+cu124 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install ninja psutil packaging && pip install flash_attn==2.7.4.post1
pip install -r requirements.txt
huggingface-cli download meituan-longcat/LongCat-Video --local-dir ./weights/LongCat-Video
Generate your first video
Text-to-video on a single GPU:
torchrun run_demo_text_to_video.py --checkpoint_dir=./weights/LongCat-Video --enable_compile
Swap the script for other modes: run_demo_image_to_video.py, run_demo_long_video.py, or run_demo_interactive_video.py. Prefer a UI? Run streamlit run ./run_streamlit.py. For multi-GPU, add --nproc_per_node=2 --context_parallel_size=2.
How LongCat-Video compares
Meituan's internal Text-to-Video MOS evaluation (higher is better):
| Overall Quality ↑ | LongCat-Video | Wan 2.2 (open) | Veo3 (proprietary) |
|---|---|---|---|
| MOS | 3.38 | 3.35 | 3.48 |
Among open-source models it edges out Wan 2.2 and lands within ~0.1 of Google's proprietary Veo3 — strong for a model you can self-host under MIT. On image-to-video it is weaker (3.17 overall, behind Seedance's 3.35).
Sources: official GitHub repo, Hugging Face model card, and the LongCat-Video Technical Report (arXiv:2510.22200). Compiled from official documentation and community reports; not independently benchmarked.