CogVideoX-5B

CogVideoX-5B arrived on 27 August 2024, three weeks after the 2B, and is the model most people mean when they say they run CogVideoX locally. It is the same shape of product — six seconds, 49 frames, 8 frames per second, 720x480 and nothing else, English prompts capped at 226 tokens — with roughly 5.6 billion parameters in the video transformer instead of 1.7, spent on picture quality rather than on longer or larger output. The one architectural difference the producer names is the positional encoding: 3D RoPE here, 3D sin-cos in the 2B. The practical threshold stayed low. In BF16 through the diffusers library, with all memory optimisations on, the producer puts inference at 5 GB of video memory, and 4.4 GB with INT8 — a desktop RTX 3060 is enough. Generation takes about 180 seconds per clip at 50 steps on an A100 and about 90 on an H100, roughly twice the 2B. Fine-tuning is where the model stops being a hobby: 63 GB per card for LoRA and 75 GB for full fine-tuning, so training needs data-centre hardware even though inference does not. The licence is the real dividing line in this family, and it runs the opposite way to what the sizes suggest. The smaller 2B was moved to Apache 2.0 on the very day this model shipped; the 5B went out under the producer's own CogVideoX Licence, which grants a copyright licence that is explicitly revocable, allows free academic use, and requires anyone using it commercially to register with the producer and stay under one million service visits per month. The producer does assign the rights to the generated content to the user, as far as local law permits.

▸News

Three generations of CogVideoX, three degrees of closing: Apache 2.0, a revocable licence, then no weights at all

9/4/2026 · Z.ai / CogVideo repository and model licences

▸Videos

No videos yet.

▸Reviews