CogVideoX splits its open license between two model sizes

The shot

Zhipu AI's CogVideoX entered its repository's own release log on 6 August 2024, when the smaller CogVideoX-2B model and an accompanying 3D Causal VAE were open-sourced; the larger CogVideoX-5B followed on 27 August 2024. A technical report, posted to arXiv on 12 August 2024, describes the model as a text-to-video diffusion system built around a 3D variational autoencoder and an "expert transformer" with expert adaptive LayerNorm, and states it can generate ten-second, 768-by-1360-pixel video at 16 frames per second.

What the documents show

The paper's claim of "state-of-the-art performance across both multiple machine metrics and human evaluations" is Zhipu AI's own reported result. The repository's release notes add a fact the paper does not: the two model sizes do not carry the same license. CogVideoX-2B's license "has been changed to the Apache 2.0 License," the notes state, while the larger CogVideoX-5B is released under a separate CogVideoX License. That document states it "allows you to freely use all open-source models in this repository for academic research," but that "users who wish to use the models for commercial purposes must register and obtain a basic commercial license," capped at "1 million visits per month" before a further license is required, and it separately restricts military or illegal use.

The workflow

For a production or tool-builder, the practical fact the license text settles is that downloading CogVideoX-5B's weights is not the same as clearing it for a monetized release: registration through the vendor's own commercial-license form is a required step the code repository does not perform automatically. CogVideoX-2B, by contrast, carries no such registration step under Apache 2.0. A human decision remains about which tier's cost, quality, and licensing terms fit a given release, and about confirming current registration requirements directly with Zhipu AI before shipping.

What the tool does not change

The license's own restriction language, including its bar on military or illegal use and its reference to Chinese law for dispute resolution, is not legal advice and does not substitute for a production's own counsel reviewing the current document, since the license itself notes it is "subject to update to a more comprehensive version." The paper's benchmark numbers likewise remain Zhipu AI's self-reported comparison until an independent evaluation states otherwise.

• Which CogVideoX model size does a planned use actually require, and does that size carry the Apache 2.0 license or the separate CogVideoX License?
• Has the production registered for a commercial license where the model's own terms require one?
• What does the current, as-retrieved version of the license state, given the license text's own note that it may be updated?

CogVideoX shows that "open-source" is not one fixed condition across a single project: two model sizes from the same repository can carry two different sets of obligations, and only the license text itself, not the paper or the marketing framing, settles which applies.