CogVideo / CogVideoX (Zhipu AI)
Source: https://github.com/zai-org/cogvideo
Structured fields
{
"archived": false,
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "Early open DiT video family with unusually explicit consumer-GPU memory tables (5B from about 5 GB BF16 with diffusers optimizations; 2B from about 4 GB FP16), LoRA fine-tuning code and DDIM inversion. A useful low-VRAM baseline and a model with a custom (non-Apache) 5B weights license.",
"source_type": "repository",
"availability": "available",
"source_owner": "zai-org (formerly THUDM)",
"next_check_at": "2026-10-31T00:00:00.000Z",
"hardware_notes": "README memory table (see evidence); measured by upstream on A100/H100; disabling optimizations can raise peak VRAM roughly 3x per the 5B card.",
"known_limitations": [
"5B fixed at 720x480 per model card, including fine-tuning.",
"Repository moved from THUDM to zai-org; the 5B card URL fetched redirected to zai-org."
],
"upstream_updated_at": "2025-11-04T11:19:04.000Z"
}Evidence claims
- [upstream_claim] README lists CogVideoX1.5-5B (1360x768, 16fps, 5/10 s), CogVideoX-5B (720x480, 8fps, 6 s) and 2B plus I2V variants. source
- [upstream_claim] README memory table: 1.5-5B from 10 GB BF16 diffusers, 5B from 5 GB BF16 / 4.4 GB INT8, 2B from 4 GB FP16 / 3.6 GB INT8, tested on A100/H100 with memory optimizations on. source
- [upstream_claim] Model card for CogVideoX-5b says 720x480 only, 49 frames, 8 fps, English prompts up to 226 tokens, fine-tuning needs ~63-75 GB per GPU. source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| code | Apache-2.0 | permitted | permitted | permitted | unknown | README states code is Apache 2.0. |
| model_weights | CogVideoX LICENSE (custom) | restricted | restricted | restricted | unknown | 5B weights: license read. Non-exclusive, non-transferable, non-sublicensable, revocable; commercial use requires registering for a free basic commercial license; commercial service limited to 1M visits/month without more licensing; PRC governing law; use restrictions. License does not explicitly address training on outputs. |
| model_weights | Apache-2.0 | permitted | permitted | permitted | unknown | README states CogVideoX-2B weights (transformer and VAE) are Apache 2.0; the 2B card was not opened. |