# CogVideo / CogVideoX — card for AI agents

Open-source text-to-video and image-to-video generation models (CogVideo, CogVideoX) with inference and fine-tuning code.

Category: AI video generators and platforms · maker THUDM · checked 2026-11-10

## Price
- Price: Free
- Free tier: Free open-source code; models can also be tried online in Hugging Face and ModelScope Spaces

## License
Apache-2.0 (code and CogVideoX-2B); CogVideoX LICENSE (CogVideoX-5B) — https://github.com/zai-org/CogVideo/blob/HEAD/LICENSE

- commercial: **conditions** (Commercial use with conditions)
  > The CogVideoX-5B model (Transformers module, include I2V and T2V) is released under
  (https://github.com/zai-org/CogVideo#model-license)
- attribution: **no** (No credit needed)
  > You must give any other recipients of the Work or Derivative Works a copy of this License
  (https://github.com/THUDM/CogVideo/blob/HEAD/LICENSE)
- in app: **conditions** (Ship inside your app with conditions)
  > The CogVideoX-5B model (Transformers module, include I2V and T2V) is released under
  (https://github.com/zai-org/CogVideo#model-license)

## Rules to follow
- CogVideoX-5B weights follow the separate CogVideoX LICENSE on Hugging Face; read it before commercial use. The code and CogVideoX-2B are Apache-2.0.
- Keep a copy of the Apache 2.0 license and the copyright and NOTICE files when you redistribute the code.
- Mark modified files with a prominent notice that you changed them.
- The README says only the CogVideoX-2B weights are Apache 2.0, so check the model license for the other weights (e.g. 5B) before commercial use.
- Do not use THUDM trade names or trademarks beyond describing where the work came from.

## How to connect
- Download only
- Repository: https://github.com/zai-org/CogVideo
- Website: https://github.com/zai-org/CogVideo

## Notes
Added automatically. License read from the official pages.

## Numbers (AgentGid data)
- Gid Score: 36
- GitHub stars: 13,214

## Lines to paste into your prompt or AGENTS.md

```
- CogVideo / CogVideoX (AI video generators and platforms). Open-source text-to-video and image-to-video generation models (CogVideo, CogVideoX) with inference and fine-tuning code. Price: free; commercial use with conditions; no credit needed; ship inside your app with conditions.
  Connect: download from https://github.com/zai-org/CogVideo (no API).
  Rule: CogVideoX-5B weights follow the separate CogVideoX LICENSE on Hugging Face; read it before commercial use. The code and CogVideoX-2B are Apache-2.0.
  Rule: Keep a copy of the Apache 2.0 license and the copyright and NOTICE files when you redistribute the code.
  Rule: Mark modified files with a prominent notice that you changed them.
  Rule: The README says only the CogVideoX-2B weights are Apache 2.0, so check the model license for the other weights (e.g. 5B) before commercial use.
  Rule: Do not use THUDM trade names or trademarks beyond describing where the work came from.
  License: https://github.com/zai-org/CogVideo/blob/HEAD/LICENSE (checked 2026-11-10; details: https://agentgid.com/assets/cogvideo/)
```

## Links
- AgentGid page: https://agentgid.com/assets/cogvideo/
- Is it free for commercial use? https://agentgid.com/assets/cogvideo/commercial-use/
- Source: https://github.com/THUDM/CogVideo/blob/HEAD/LICENSE
- Source: https://github.com/THUDM/CogVideo


## CogVideoX: short AI video clips on your own machine

CogVideoX is a family of free AI models that make short video clips from a text description or from a still picture. You run it yourself on a computer or rented server with a strong graphics card, so no outside video service is needed. The small model is free for any use; the larger ones need a free sign-up before commercial use.

### Where to use it
- A clip from a text description: Describe a scene in English, such as a panda playing guitar in a forest. You get a clip of about six seconds to use as a draft or a short loop.
- Bring a still picture to life: Start from your own photo or drawing and add a short description of the motion; the picture becomes the first frame of the clip.
- Turn a short idea into a full prompt: The models work best with long, detailed descriptions. A helper script lets a chat AI expand a one-line idea into such a description first.
- Teach it your own style: Train the model a little further on your own short clips, for example product shots or interior scenes, so new clips follow that look.
- Moving backgrounds for a landing page: Make a few calm, wordless clips, such as clouds or water, and play them muted behind a page header instead of buying stock footage. (idea, not from the official pages)
- Quick previews for a video script: Turn each scene of an ad or explainer script into a rough clip, so the team can judge the order and mood before real filming. (idea, not from the official pages)

### How to start
- Use a machine with an NVIDIA graphics card and Python 3.10 to 3.12, then install the libraries the model card lists.

```
pip install --upgrade transformers accelerate diffusers imageio-ffmpeg
```

- Load the model by its name and give it a long English description; the first run downloads the model files.

```
pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-5b",
    torch_dtype=torch.bfloat16
)
```

- No graphics card at hand? The project links ready notebooks that run on free Google Colab T4 machines.

### License
The code and the small 2B model are Apache 2.0, free for paid products. The 5B models use their own license, which is free for research. Commercial use needs a free sign-up and is capped at one million visits a month.

### Good to know
- You need a strong NVIDIA graphics card or a rented one. One clip takes minutes, not seconds, even on top data-center cards.
- Write prompts in English. Other languages should be translated first, for example by a chat AI.
- Clips are short and small. The first models give about six seconds at 720 by 480; version 1.5 gives five or ten seconds at a higher size.
- The 5B models may not be used for military or illegal purposes, and their license can be updated later.
- The project now lives under the zai-org account; old THUDM links forward there. The newest models were released in November 2024.

Checked 2026-11-10 against: https://github.com/THUDM/CogVideo, https://github.com/zai-org/CogVideo, https://github.com/zai-org/CogVideo/blob/HEAD/README.md, https://github.com/zai-org/CogVideo/blob/HEAD/LICENSE, https://github.com/zai-org/CogVideo/blob/HEAD/MODEL_LICENSE, https://github.com/zai-org/CogVideo/blob/HEAD/requirements.txt, https://github.com/zai-org/CogVideo/blob/HEAD/inference/convert_demo.py, https://huggingface.co/THUDM/CogVideoX-5b/raw/main/README.md, https://huggingface.co/THUDM/CogVideoX-5b/raw/main/LICENSE, https://huggingface.co/THUDM/CogVideoX-5b-I2V/raw/main/README.md, https://huggingface.co/THUDM/CogVideoX-2b/raw/main/README.md, https://huggingface.co/docs/diffusers/api/pipelines/cogvideox, https://pypi.org/pypi/diffusers/0.42.1/json

### Mistakes to avoid
- Don't: Asking for any number of frames, such as 60 → Do: Use 8N+1 frames (default 49) for CogVideoX-2B/5B and 16N+1 (default 81) for CogVideoX1.5 (The model table fixes the frame counts per model family.)
- Don't: Passing a short or non-English prompt straight to the model → Do: Expand the idea into a long English description first (convert_demo.py or any chat model) (The models were trained on long English prompts.)
- Don't: Keeping enable_sequential_cpu_offload() on when running on several GPUs → Do: Turn that option off for multi-GPU runs; keep it for single small GPUs (The README says that optimization must be disabled for multi-GPU inference.)
- Don't: Running the 2B model in BF16 like the 5B models → Do: Use torch.float16 for CogVideoX-2B and torch.bfloat16 for the 5B models (Each model should run in the precision it was trained in.)
- Don't: Shipping a commercial product on CogVideoX-5B because "the code is Apache 2.0" → Do: Use CogVideoX-2B for Apache 2.0 terms, or register for the 5B commercial license first (Only the code and the 2B weights are Apache 2.0; the 5B weights have their own license.)

### Questions
- Can I use CogVideoX in a paid product? The 2B model and the code, yes, under Apache 2.0. The 5B models need a free commercial sign-up, and above one million visits a month you must contact the team.
- Who owns the videos I make? The license says the rights to generated content belong to the user, as far as local law allows.
- Can I try it without my own graphics card? Yes. There is an online demo on Hugging Face, and the project links Colab notebooks that run on free T4 machines.
- Why do short prompts give poor clips? The models were trained on long descriptions, so the team suggests expanding a short idea with a chat AI first.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
