LongCat Video 2026: local setup, parameters, browser fallback
Distilled or regular checkpoints, prompt and parameter settings, ComfyUI memory fixes — or run LongCat in the browser with no install.
Browser fallback: if a local install or a ComfyUI run fails — CUDA errors, an out-of-memory stop, or a checkpoint that will not load — generate with the hosted LongCat image-to-video workflow instead. There is nothing to install and the credit cost of the run is quoted before you submit it.
What is LongCat Video?
LongCat Video (also written Long Cat Video or Longcat-Video) is Meituan's open-source AI video generation model, released on 25 October 2025. It is a single dense 13.6-billion-parameter model that handles text-to-video, image-to-video and video continuation in one framework, which is why it keeps characters, colour and camera behaviour stable across longer shots.
The model weights are released under the MIT License, so you may run them locally, modify them and use the output commercially without a subscription. What is not free is compute: a local run costs you your own GPU time and electricity, and the hosted workspace on this site is metered in credits. With the MIT license you can:
- Run the model locally on your own hardware, with no per-clip fee
- Modify or fine-tune the weights for your own use
- Use the results commercially under the MIT terms
- Generate here in the browser without installing anything, paying per run in credits
"Unlimited" is the wrong word for either route. Locally, your hardware is the ceiling — every extra minute of video is another inference job. On the hosted side there is no free daily allowance and no unlimited plan: a new account starts with 150 one-time signup credits and paid plans are credit packs.
Key features of LongCat Video
| Feature | Specification |
|---|---|
| Frames per clip | 17-961 here; 961 frames ≈ 32s at 30fps, ≈ 64s at 15fps |
| Resolution | 480p and 720p output |
| Workflows | Text-to-video, image-to-video, continuation in one model |
| License | MIT for the model weights (no Meituan trademark or patent rights granted) |
| Parameters | 13.6 billion, dense transformer |
| Avatar model | LongCat-Video-Avatar-1.5 (May 2026) is a separate audio-driven model |
LongCat Video vs Sora vs Veo 3: what we will not print
Posts like this one usually carry a table of competitor prices and star ratings. We do not print one, because we cannot verify what Sora, Veo, Kling, Runway or Pika charge today from an official source, and a stale guess is worse than no table.
What we can state comes straight from Meituan's repository and technical report:
- One dense 13.6B model covers text-to-video, image-to-video and video continuation; the weights are MIT licensed.
- Upstream describes 720p, 30fps generation through a coarse-to-fine strategy along both the temporal and spatial axes, with Block Sparse Attention for the long-sequence passes.
- Meituan's own internal MOS benchmark for text-to-video scores LongCat-Video at 3.38 overall quality against 3.48 for Veo3 and 3.35 for Wan 2.2-T2V-A14B. That is the model authors' own benchmark, not an independent ranking — treat it as a claim to test.
- LongCat-Video-Avatar-1.5 (May 2026) adds audio-driven talking heads with a Whisper-Large-v3 audio encoder.
The honest comparison method is the only one that holds: generate the same prompt on each tool and watch the output. Your own footage is the benchmark that matters.
How to Use LongCat AI Video Generator
Option 1: Use Our Online Platform (Easiest)
The fastest way to start with LongCat Video AI is through our web platform:
- Visit longcat-video.org
- Create an account — new accounts start with 150 one-time signup credits, valid 30 days, enough for one Distilled 480p run
- Upload a reference image (or start from a text prompt in text-to-video mode)
- Read the quoted credit cost for the run, then click "Generate Video"
- Download the result
No installation required — it runs in the browser. Every run costs 150-300 credits depending on model and resolution, and the exact figure is shown before submission. See what a run costs.
Option 2: LongCat Video download (local installation)
To run the official weights yourself, follow Meituan's published path:
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
# PyTorch build for your own CUDA version (upstream example: CUDA 12.4)
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
# weights
pip install "huggingface_hub[cli]"
huggingface-cli download meituan-longcat/LongCat-Video --local-dir ./weights/LongCat-Video
# text to video
torchrun run_demo_text_to_video.py --checkpoint_dir=./weights/LongCat-Video --enable_compileSwap in run_demo_image_to_video.py, run_demo_video_continuation.py or run_demo_long_video.py for the other modes, and add --nproc_per_node=2 --context_parallel_size=2 for a two-GPU run.
What upstream actually publishes about requirements:
- The install path assumes an NVIDIA GPU with a recent CUDA toolchain (the example pins CUDA 12.4). No macOS path is documented.
- Meituan publishes no universal VRAM, RAM or disk minimum, and no per-GPU speed table. Check the Hugging Face model card for the exact checkpoint size before you commit disk space.
- Any "RTX 4090 = 3 minutes per clip" style table you see elsewhere, including older versions of this page, is somebody's measurement on their machine, not a specification. Measure on your own hardware.
Option 3: LongCat Video ComfyUI workflow
ComfyUI supports LongCat through the community rather than an official Meituan extension — there is no ComfyUI repository in the meituan-longcat GitHub organisation, so treat any tutorial that clones one as outdated.
- Update ComfyUI and install ComfyUI Manager so extensions resolve their own requirements.
- Add the wrapper that carries the LongCat nodes: kijai/ComfyUI-WanVideoWrapper. Our own PainterLongVideo tutorial and avatar guide pin the same wrapper; record the commit hash so a node rename cannot break every workflow at once.
- Download Comfy-format LongCat weights from the community LongCat-Video_comfy repository, or convert the official checkpoints.
- Load an example workflow from the wrapper instead of an untraceable JSON from a video description, then run one known-good input before you change anything.
Related workflows: the PainterLongVideo tutorial documents the motion controls, and LongCat-Video-Avatar-1.5 is the audio-driven path.
LongCat Video Prompt Engineering
Getting great results from LongCat AI Video Generator requires understanding prompt engineering. Here's the complete guide:
The 5-Component Prompt Structure
Every effective LongCat Video prompt should include:
- Scene Description — Visual elements, setting, atmosphere
- Motion Direction — How objects/characters move
- Cinematographic Elements — Camera movement, lighting, perspective
- Style References — Visual aesthetics (photorealistic, anime, documentary)
- Technical Qualifiers — Resolution and quality indicators
Example: Weak vs Strong Prompts
❌ Weak Prompt:
a car driving on a road✅ Strong Prompt:
A sleek red sports car driving down a winding coastal highway at sunset.
The camera follows alongside the vehicle, capturing reflections of the
golden sun on its polished surface. The scene transitions from close-up
details of the wheels to a wide aerial shot revealing the dramatic
coastline below. Cinematic lighting, photorealistic, 4K quality.Motion Vocabulary for LongCat Video
Use specific motion terminology for better results:
| Category | Words to Use |
|---|---|
| Verbs | floating, accelerating, dissolving, emerging, circling, transforming |
| Adverbs | smoothly, gradually, rapidly, rhythmically, gently, dramatically |
| Transitions | transforming into, fading to, zooming out to reveal, cutting to |
| Camera | pan, tilt, dolly, crane shot, tracking shot, aerial view |
Negative prompts
Negative prompts are a community habit, not an upstream requirement — Meituan's repository documents no mandatory negative prompt for LongCat-Video. A list in this style is what most LongCat and Wan workflows start from; keep the artifact half (blur, text, extra limbs) and delete anything that contradicts your own shot:
Bright tones, overexposed, static, blurred details, subtitles,
style, works, paintings, images, overall gray, worst quality,
low quality, JPEG compression residue, ugly, incomplete, extra fingers,
poorly drawn hands, poorly drawn faces, deformed, disfigured,
misshapen limbs, fused fingers, still picture, messy backgroundLongCat Video parameters guide
These are the ranges and defaults of this site's generator, so the numbers here match what the form accepts:
| Parameter | Range | Default and practical guidance |
|---|---|---|
| Resolution | 480p / 720p | 720p pairs with 30fps, 480p with 15fps |
| num_frames | 17-961 | Default clip is 162 frames; 900@15fps ≈ 60s, 961@15fps ≈ 64s |
| num_inference_steps | 8-50 | Default 40 |
| guidance_scale | 1-10 | Default 4; higher stays closer to the prompt |
| fps | 1-60 | 30 for 720p, 15 for 480p unless you have a reason to change |
Two settings bite people who ignore them. Distilled checkpoints on this site are capped at 81 frames — ask for more and the run is automatically upgraded to the regular model and billed at its higher credit cost. And a longer clip is not a longer film: continuation means repeating the generation, so inspect the join between segments before you trust a minute-long run.
LongCat Video costs: credits, not tiers
There is no "Free tier / Pro tier / Enterprise" ladder on this site. One credit system covers every model, and the exact cost of a run is quoted before you press generate.
Cost of one run
| Run | Credits |
|---|---|
| LongCat Video, Distilled 480p | 150 |
| LongCat Video, regular 480p | 200 |
| LongCat Video, Distilled 720p | 250 |
| LongCat Video, regular 720p | 300 |
| Image generation or edit | 80 |
| Music track | 100 |
A new account starts with 150 one-time signup credits, valid for 30 days — exactly one Distilled 480p video. That is the whole free allowance; there is no daily free quota, and no paid plan is unlimited.
Plans (every price and credit amount below is the live catalog on /pricing):
| Plan | Price | Credits | Validity |
|---|---|---|---|
| Try monthly | $4.9 / month | 1,000 | 1 month |
| Pro monthly | $49 / month | 8,000 | 1 month |
| Max monthly | $149 / month | 30,000 | 1 month |
| Try annual | $5.75 / month | 15,000 | 1 year, granted upfront |
| Pro annual | $29 / month | 96,000 | 1 year, granted upfront |
| Max annual | $83.25 / month | 360,000 | 1 year, granted upfront |
| Try once | $9.9 | 500 | forever |
| Pro once | $89 | 5,000 | forever |
| Max once | $389 | 50,000 | forever |
Which one to pick, by finished runs: Try monthly (1,000 credits) covers six Distilled 480p runs or three regular 720p runs. Pro monthly (8,000) covers 26 regular 720p runs. Max monthly (30,000) covers 100 of them. Long video is a continuation job, so budget several runs per finished minute — a 60-second result is not one 60-second generation.
Self-hosting is the other cost model: the weights are MIT licensed and free to download, and you pay in GPU time instead. Neither route is cheaper in the abstract; a hosted plan removes the hardware, a local install removes the per-run charge.
LongCat Video open source on GitHub
Meituan LongCat Video is open source. The official sources:
- Model repository: github.com/meituan-longcat/LongCat-Video
- Model weights: huggingface.co/meituan-longcat/LongCat-Video
- Avatar 1.5 weights: huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.5
- Technical report: arXiv 2510.22200
There is no official Meituan ComfyUI node in that organisation; the ComfyUI route goes through the community wrapper linked above.
Contributing to LongCat Video
# Fork and clone
git clone https://github.com/YOUR_USERNAME/LongCat-Video
cd LongCat-Video
# Create a branch
git checkout -b feature/your-feature
# Make changes and submit PR
git push origin feature/your-featureFrequently Asked Questions
Is LongCat Video really free?
The model weights are MIT licensed, so running them yourself costs nothing beyond your own hardware and electricity, and commercial use is allowed. The hosted workspace here is a paid service: a new account gets 150 one-time signup credits (one Distilled 480p run), after which plans start at $4.9/month for 1,000 credits. Read the cost breakdown before you plan a long job.
What's the difference between LongCat Video and LongCat Image?
- LongCat Video — video generation: text-to-video, image-to-video, continuation (13.6B parameters, MIT weights)
- LongCat Image — image generation and editing (6B parameters, Apache-2.0)
They are separate models with different architectures, both published by Meituan.
Can I run LongCat Video on my laptop?
Only if the laptop has a supported NVIDIA GPU — the official install path is built around CUDA (the upstream example pins CUDA 12.4) and Meituan publishes no macOS route. There is also no published universal VRAM figure: what fits depends on checkpoint format, resolution, frame count and attention backend, so measure with a three-second test clip on your own machine instead of trusting a table. If that is not your setup, the browser generator needs nothing installed.
How does LongCat Video compare to Kling, Runway, and Pika?
We do not print a price or star-rating table for those tools: their plans and limits change and we cannot verify current numbers from an official source. The comparable facts about LongCat Video are that its weights are open (MIT), it runs locally, and one model covers text-to-video, image-to-video and continuation. Test the tools you are choosing between with the same prompt and the same length.
What is LongCat Distilled?
Distilled LongCat checkpoints trade a little quality for faster, cheaper inference. On this site a Distilled run costs 150 credits at 480p and 250 at 720p, against 200 and 300 for the regular model, and Distilled checkpoints are capped at 81 frames — exceed that and the job is automatically upgraded to the regular model at its higher price. Use Distilled for tests and short clips; switch to the regular model for the final long take.
Does the hosted LongCat tool make talking avatars?
No. The generator on this site takes a still image plus a motion prompt and returns a clip; it does not accept an audio file and does not produce lip-synced speech. Audio-driven talking heads are a different model — LongCat-Video-Avatar-1.5, which you run yourself. The LongCat video avatar guide covers the checkpoint, the inputs worth preparing and the honest VRAM limits, and our ComfyUI avatar tutorial walks the local setup.
How long can one LongCat clip be?
Between 17 and 961 frames in this site's generator, which is about 32 seconds at 30fps or 64 seconds at 15fps. Anything longer is continuation: the same model generates a further segment from your existing footage, so plan for several runs per finished minute and inspect each join for identity and colour drift.
LongCat Video Tutorials
Learn more with our detailed tutorials:
Getting Started
- Complete Guide to LongCat Video
- LongCat video avatar: the audio-driven path
- PainterLongVideo ComfyUI tutorial
ComfyUI Workflows
Advanced Topics
Multi-Language Guides
Conclusion
LongCat Video is a rare case of a frontier-class video model published with open weights: one 13.6B model for text-to-video, image-to-video and continuation, MIT licensed, and documented well enough to reproduce locally.
Two things decide whether it works for you. The first is compute — a local run is free of fees but not of hardware, and the hosted route is metered in credits at a rate you can see before every job. The second is honesty about limits: no model turns a hard cut into a film, and continuation still needs a human eye on the joins.
Ready to start? Generate your first clip in the browser with your 150 signup credits, or compare the credit packs on the pricing page.
Last updated: January 25, 2026
Keywords: longcat video, long cat video, longcat video ai, long cat video ai, longcat-video, longcat ai video generator, longcat ai video, long cut video, longcat video generator, longcat video free, longcat video download, meituan longcat video, longcat video github, longcat video comfyui, longcat video open source, longcat distilled