LongCat Video Complete Guide: Setup and Costs

What LongCat Video actually is

LongCat Video is a dense 13.6-billion-parameter video model from Meituan's LongCat team, published on 25 October 2025. The parameter count is not the interesting part. The interesting part is the packaging: one model unifies text-to-video, image-to-video and video continuation, instead of asking you to install three different systems with three different quality signatures.

Continuation is what changes a real workflow. The model is pretrained natively on the job of carrying motion forward from the frames it has just produced, which is why the project talks about minute-scale output rather than a hard stop after a few seconds. In an editing pipeline that means you can generate a section, inspect the join, and extend rather than throw the clip away and re-roll it from scratch.

The weights are public. The repository's license agreement states that the model weights are released under the MIT License. The same agreement grants no Meituan trademark or patent rights, so treat the release as reusable technology rather than an officially supported Meituan product with a support contract behind it.

The upstream repository documents 720p output at 30 frames per second, a coarse-to-fine strategy applied along both the temporal and the spatial axes, Block Sparse Attention, multi-reward GRPO reinforcement learning from human feedback, and native pretraining on continuation, which the authors describe as yielding minutes-long videos without color drifting or quality degradation. Those are the authors' claims about their own system; we report them as claims, because we have not reproduced them independently.

There is exactly one quantitative comparison published upstream, and it deserves careful labelling: an internal MOS (mean opinion score) test for text-to-video, run by the model's own authors. LongCat Video scored 3.38 overall, against 3.48 for Veo3, 3.36 for PixVerse-V5 and 3.35 for Wan 2.2-T2V-A14B. It is a vendor-run benchmark, not an independent ranking, and the fair reading is narrow: a 13.6B open-weight release landed in the same band as the commercial systems it was measured against.

One thing to separate early, because it is the single most common confusion around this family: LongCat-Video-Avatar is a different model with a different job. It is audio-driven, not prompt-driven. The original Avatar release came on 16 December 2025 with a wav2vec2 audio encoder. Avatar-1.5, released 21 May 2026, replaces Wav2Vec2 with Whisper-Large-v3 for more accurate lip sync, accepts single- and multi-stream audio, switches between 480P and 720P through the --resolution flag, supports INT8 DiT quantization with --use_int8, and requires distillation sampling through --use_distill, which brings inference down to 8 steps. The technical report is at arXiv 2510.22200 if you want the architecture in detail.

Three ways to run LongCat Video

You do not have to pick one path forever, but you should pick one to start, because each has a different failure mode.

1. The hosted generator on this site

The browser tool here is built around a still image plus a motion prompt: you upload a picture, describe how it should move, choose a length and a resolution, and pay in credits per job. Nothing to install, no CUDA version to argue with, and the credit cost of the run is shown before you submit it. This is the fastest way to find out whether your prompt idea works at all before you commit to a local install.

2. The official local install

Upstream publishes a specific, reproducible route. Run it in order; the PyTorch line has to match your CUDA build rather than the exact example below.

git clone --single-branch --branch main https://github.com/meituan-longcat/LongCat-Video
conda create -n longcat-video python=3.10
# install the PyTorch build that matches your CUDA (upstream example for 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
pip install "huggingface_hub[cli]"
huggingface-cli download meituan-longcat/LongCat-Video --local-dir ./weights/LongCat-Video

Then run the demo that matches the task:

torchrun run_demo_text_to_video.py --checkpoint_dir=./weights/LongCat-Video --enable_compile
# other entry points: run_demo_image_to_video.py, run_demo_video_continuation.py,
# run_demo_long_video.py, run_demo_interactive_video.py
# multi-GPU adds: --nproc_per_node=2 --context_parallel_size=2

What upstream does not publish matters as much as what it does. There is no universal VRAM, RAM or disk minimum, and no per-GPU speed table. If a blog post hands you a neat table of minutes per clip on a named GPU, that table came from the author's machine, not from Meituan. Check the Hugging Face model card for the exact checkpoint size before you clear disk space for it.

3. The ComfyUI community route

ComfyUI supports LongCat Video through the community, not through Meituan. The meituan-longcat GitHub organisation publishes no ComfyUI repository, so anyone who tells you there is an official Meituan ComfyUI node is wrong, and any install command that clones a LongCat node from that organisation will fail. What does exist is a working community pair: Comfy-format weights at Kijai/LongCat-Video_comfy, and the LongCat nodes that ship inside kijai/ComfyUI-WanVideoWrapper. If you are starting from a bare ComfyUI, install ComfyUI first, then the wrapper, then download the Comfy-format weights into your models folders. This path is also the one our own avatar and PainterLongVideo guides are pinned to.

The parameters that actually matter

Four controls decide almost everything about a run. Frames and fps set how long the clip is and how it feels; inference steps and guidance set how refined and how prompt-faithful it is.

Frames are the real duration dial. The default short clip is 162 frames, which is roughly 10.8 seconds at 15fps and roughly 5.4 seconds at 30fps. Minute-scale examples sit near the ceiling: 900 frames at 15fps is about 60 seconds, 961 frames at 15fps is about 64 seconds, and 961 frames at 30fps is about 32 seconds. Choose fps first if the destination platform demands it, then set frames to hit the length you need, because asking for 30fps at the top of the frame range buys you a shorter clip than 15fps does.

Inference steps trade compute for refinement, and guidance scale trades prompt obedience for naturalness: higher guidance pulls the output closer to your written prompt. If a shot drifts away from what you described, raise guidance before you rewrite the prompt from scratch.

There is one hard limit to know on the hosted tool: distilled runs are capped at 81 frames. Ask for more than that with a distilled model selected and the request is automatically upgraded to the regular model, which is why a long distilled request bills at the regular rate.

Writing prompts that survive motion

A single sentence is not a prompt, it is a wish. The structure that holds up across text-to-video, image-to-video and continuation is five parts, in this order: subject and action, environment, camera, lighting, and style or technical finish. Each part removes one degree of freedom the model would otherwise guess at.

Weak: "a woman presenting in an office".

Strong: "A product manager walks to a glass wall and points at a chart, in a modern conference room with desks visible behind her, slow push-in from waist height, soft daylight from the left plus warm interior lamps, corporate documentary look, sharp focus, 720p, 30fps".

The second prompt is not longer for the sake of length. It names the action verb, fixes the camera move, and states the light direction, so the model does not have to invent a scene change it cannot undo.

Motion vocabulary is a genuine skill, so build a small personal list. Useful verbs: walk, turn, lift, pour, open, drift, unfold, settle. Useful adverbs: slowly, steadily, gently, sharply. Useful transitions: cut to, dissolve into, camera follows, time-lapse into. Camera language is its own vocabulary and the model reads it: "camera pushes in" for a smooth move closer, "camera pulls back" to reveal context, "pan across the scene" for a horizontal sweep, "tracking shot" to follow a subject, "drone rises" for an upward reveal, and "locked-off static shot" when you want the frame to stop moving entirely.

For image-to-video, the still carries the composition, so prompt the motion you want added and nothing else. Describing a new environment in an image-to-video prompt fights the input frame. For continuation, describe where the story goes next and keep lighting direction consistent with the previous segment, otherwise the join becomes a visible cut with a lighting flip attached to it.

What a run costs

Every model on this site shares one credit price list, and the exact cost of a job is displayed before you submit it.

New accounts receive 150 one-time signup credits, valid for 30 days. In the code that grant is described as one-time signup credits for a Distilled 480p video, so the honest framing is: the signup credits cover exactly one short distilled 480p clip, which is enough to test an idea and judge whether the tool fits your work.

After that, credits come in three shapes. Annual plans grant all credits upfront and those credits are valid for one year; monthly plans reset on a one-month validity window; one-time packs never expire.

To turn that into a mental model: a one-time Try pack at $9.9 buys 500 credits, which is three distilled 480p runs, one regular 720p run plus change, or 100 credits short of three regular 480p runs. A monthly Try plan at $4.9 covers six distilled 480p runs, or five regular 480p runs. The plans, the exact credit maths per model, and the refund policy all live on /pricing; read that page rather than trusting a summary, including this one.

Honest limits

There is no published universal hardware minimum. Upstream gives no VRAM, RAM or disk floor and no speed table, so any single VRAM number quoted as "enough" is somebody's machine, not a specification. Check the model card for the checkpoint size.

Distilled is not free of limits. On this site distilled runs are capped at 81 frames and longer requests upgrade to the regular model. If you plan a minute-long shot, price it as a regular run from the start.

Continuation joins need to be watched. LongCat Video is pretrained for continuation and the authors describe it as drift-free over minutes, but any multi-segment assembly has a seam at each join. Generate, watch the join at full speed and frame by frame, and only then extend again. Do not assume a seam is invisible because the feature is called continuation.

The hosted tool does not do talking heads. A run here takes a still image and a motion prompt. It does not accept audio input and it does not produce lip-synced speech. If you need an audio-driven avatar, that is the Avatar-1.5 setup, described in our LongCat Video Avatar guide, and the audio work happens in that local pipeline rather than in the browser.

We do not publish vendor comparison tables. We cannot verify other vendors' current pricing, plan limits or duration caps, so a table scoring LongCat Video against Runway, HeyGen, Synthesia or Pika would be stale the week it was written and would not be honest. What is verifiable is above: the MIT-licensed weights, the documented 720p/30fps capability, the parameter and credit figures on this page, and the model authors' own MOS test labelled as theirs.

FAQ

Is LongCat Video free?

The weights are, and that is the part people usually mean: the model weights are released under the MIT License, so you can download, run and build on them. The hosted generator on this site is not a free-forever service. New accounts get 150 one-time signup credits valid for 30 days, which is one short distilled 480p video, and after that you buy credit packs. There is no free daily allowance, and no plan that removes the per-run credit cost.

Can I run it locally?

Yes, and upstream documents the exact commands. You need a CUDA-capable GPU machine and a PyTorch build that matches your CUDA version. Follow the install sequence in this guide, download the checkpoint with huggingface-cli, and start with run_demo_image_to_video.py if your first goal is animating a still.

Which checkpoint should I download?

The official weights are at meituan-longcat/LongCat-Video. Upstream does not publish a universal checkpoint size in a form we can quote as a requirement, so check the model card on Hugging Face and size your disk from that. If you would rather work in ComfyUI, use the Comfy-format weights at Kijai/LongCat-Video_comfy with the nodes from ComfyUI-WanVideoWrapper.

How much VRAM do I need?

Upstream publishes no universal VRAM minimum, so anyone quoting a single number is quoting their own card. What you can control on the hosted side is the run: 480p is cheaper and lighter than 720p, distilled is cheaper and lighter than regular, and distilled runs are capped at 81 frames here. If you are sizing local hardware, treat the model card as the source of truth and expect to test.

Does the hosted tool make talking avatars?

No. The browser generator takes a still image plus a motion prompt, has no audio input, and cannot produce lip-synced speech. Audio-driven avatars require LongCat-Video-Avatar, and Avatar-1.5 specifically needs the local setup covered in our LongCat Video Avatar page.

What does a single run cost?

150 credits for Distilled 480p, 200 for regular 480p, 250 for Distilled 720p and 300 for regular 720p. Your signup grant of 150 one-time credits therefore covers one short distilled 480p clip. Longer distilled requests are upgraded to the regular model automatically, so a long request bills at the regular rate.

A realistic learning path

Start narrow. Pick one still image you care about and generate a single short distilled 480p clip at the default length. The point of the first run is not the output, it is learning how the parameters feel: what 15fps versus 30fps does to the sense of speed, and how much a guidance change alters prompt obedience.

Then build prompt discipline. Rewrite five prompts into the five-part structure, keep a file of motion verbs and camera terms that worked, and compare the same prompt at two guidance values. Most people improve faster here than they do by collecting workflows.

Only then move to length. Once a clip at the default length is reliably what you intended, test continuation and inspect each join. From there, choose your production path: stay in the browser if you want volume without maintenance, or install locally if you need the full checkpoint under your own control. If your work is genuinely about talking humans, go straight to the Avatar track instead of fighting the hosted tool into doing something it does not do.

Where to go next

• Open the image-to-video generator and price a clip with your own image.
• Compare credit plans to work out whether one-time, monthly or annual credits fit your usage.
• LongCat Video Avatar for the audio-driven, lip-synced pipeline the browser tool does not cover.
• LongCat Avatar ComfyUI tutorial if you want to build the avatar workflow inside ComfyUI.
• PainterLongVideo ComfyUI tutorial for long-form work in the same community ecosystem.