LongCat-Video: Meituan video model, 7 demo modes

TL;DR: LongCat-Video is a Python model repository from the Meituan LongCat organization, MIT licensed, with 8,758 stars. The GitHub description is empty, but the repository ships seven demo scripts whose names map cleanly onto video generation modes: text, image, continuation, long form, interactive, and two avatar audio-driven variants. It is pure Python code at roughly 512 KB inside a 1.3 GB repository.

The file names are the spec

Since the description field is blank, the top-level tree carries the argument. Every demo script states its mode in the filename:

Script	Mode
run_demo_text_to_video.py	prompt to video
run_demo_image_to_video.py	still image to video
run_demo_video_continuation.py	extend an existing clip
run_demo_long_video.py	longer generation runs
run_demo_interactive_video.py	interactive session
run_demo_avatar_single_audio_to_video.py	one audio track driving an avatar
run_demo_avatar_multi_audio_to_video.py	multiple audio tracks
run_streamlit.py	browser UI over the model

Continuation and long-form generation are the two entries that separate this from a typical text-to-video drop. Continuation means feeding a clip back in and producing the next segment, which is how you build sequences longer than a single generation window. Long video is the same problem approached from the other side.

The interactive script is the one worth a second look, since a Streamlit front end over a video model implies steering during generation rather than only setting a prompt up front.

One language, one package

The language breakdown is unambiguous. There is no inference service, no frontend build, no compiled extension:

{
  "Python": 523907
}

The implementation sits in a single longcat_video package directory. For a repository of this size that is a compact surface, and it means you can read the whole model wrapper without wading through infrastructure.

Why a 1.3 GB repository holds 512 KB of code

The API reports the repository at 1,366,898 KB, which is about 1.3 GB, against 523,907 bytes of Python. Those two numbers do not sit together comfortably, and the difference is explained by the assets directory at the root. Demo media, sample inputs, and example outputs are committed to the tree.

The arithmetic makes the split obvious:

repo=meituan-longcat/LongCat-Video
size_kb=$(curl -sL "https://api.github.com/repos/$repo" | jq .size)
py_bytes=$(curl -sL "https://api.github.com/repos/$repo/languages" | jq .Python)
echo "repo $((size_kb / 1024)) MB total, python $py_bytes bytes"

Expect a slow clone for that reason, and do not assume the repository size reflects model weights. Where the weights actually live is a question the metadata does not answer.

The avatar path has its own dependencies

Alongside requirements.txt there is a separate requirements_avatar.txt. Splitting dependency sets this way usually means the avatar feature pulls in extra packages, most likely audio or face-related ones, that the text and image paths do not need.

That is a useful signal if you only want the core generation modes. Install the base requirements, run the text, image, or continuation demo, and leave the heavier avatar stack alone until you know you need it.

The rest of the metadata

Field	Value
Full name	meituan-longcat/LongCat-Video
Homepage	none declared
Language	Python
License	MIT
Stars	8,758
Forks	1,526
Open issues	80
Created	2025-10-25
Last push	2026-05-27

The fork ratio is high at roughly one fork per six stars, consistent with a model people download and run locally. The repository was created on 2025-10-25 and last pushed on 2026-05-27, about four months before this writing, so treat version-specific claims with care and pin your environment.

When it fits

Worth a look if you are evaluating open video generation and want several input modes from one codebase rather than a single text-to-video script. Continuation support in particular is the part that most affects whether you can build anything longer than a few seconds.

MIT licensing means the code is open to reuse with no restriction beyond attribution and notice, and the small Python surface makes that reuse practical.

When to skip it

Skip it if you need a hosted API rather than a repository. Nothing in the metadata points to an inference service, and running a video model on your own hardware is a different commitment from calling an endpoint.

Also read the license and the asset situation before you build on it. MIT covers the code, while the bundled media in that 1.3 GB has its own provenance, and a model release with no declared homepage leaves questions about intended use that you answer by reading the repository rather than a card.

Source and Accuracy Notes

• Project page: https://github.com/meituan-longcat/LongCat-Video
• Source repository: meituan-longcat/LongCat-Video (verified via GitHub API html_url)
• License: MIT (verified via GitHub API license.spdx_id)
• Stars at time of writing: 8,758
• Additional verification: languages and contents API endpoints