CogVideoX
Open text-to-video and image-to-video models from Zhipu AI, with 2B and 5B sizes you run yourself.
CogVideoX offers 2B and 5B video models with diffusers and SAT code.
CogVideoX is a family of open text-to-video and image-to-video models with the code to run and fine-tune them. It suits developers and technical creators who want local video generation on a modest NVIDIA card.
What it is
The repo ships inference and fine-tuning code for CogVideoX-2B, CogVideoX-5B and the 1.5 releases, in both diffusers and the SAT toolkit. The README tabulates VRAM use per model and per precision.
Who it's for
It suits people comfortable with Python who want an older but documented open model. It does not suit anyone who wants a point-and-click app, and the 5B weights carry a separate licence from the code.
Key features
2B and 5B models
Two model sizes, plus I2V variants, trade quality against memory.
Diffusers support
Run the models through the diffusers library with memory optimisations.
SAT toolkit
A second inference and training path using the SAT framework.
INT8 options
The README lists INT8 runs through torchao that trim VRAM further.
Fine-tuning code
Scripts for LoRA and full fine-tuning ship in the repository.
VRAM table
Memory needs are listed per model, precision and multi-GPU setup.
Up and running in 3 steps
Our honest verdict
What's great
- The 2B model is listed from about 4GB of VRAM with diffusers optimisations, far below most video models.
- Code is Apache-2.0 and the CogVideoX-2B weights are Apache-2.0 as well.
- Both diffusers and SAT paths are documented, with fine-tuning scripts included.
- The README gives per-model VRAM figures, so you can size a GPU before downloading weights.
Where it falls short
- You need Python, a CUDA GPU and several gigabytes of weight downloads; there is no installer or web app.
- The CogVideoX-5B weights sit under a separate CogVideoX licence, so read it before commercial use.
- The latest news item in the README is from March 2025, so newer models from other labs may beat it.
- The SAT path lists far higher VRAM figures than the diffusers path, so pick the diffusers route on small cards.
Pricing & license
Free under Apache-2.0 for the code. The only cost is the GPU and storage you run it on; the repo shows no paid hosted plan.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
How much VRAM does CogVideoX need?
The README lists about 4GB for the 2B model and about 5GB for the 5B image-to-video model with diffusers optimisations on. Turning optimisations off can multiply peak memory.
Can I use CogVideoX commercially?
The code is Apache-2.0 and the 2B weights are Apache-2.0. The 5B weights use a separate CogVideoX licence, so check its terms first.
Does CogVideoX have a web interface?
The repo is code and scripts. It runs from the command line or through libraries such as diffusers, and the README describes no hosted app.
