Open weightsZhipu AI (THUDM / Tsinghua)/10 trustVerified

CogVideoX-5B

Name: CogVideoX-5B
Rating: 6

Zhipu AI and Tsinghua's 5B open-weights text-to-video diffusion model with an expert transformer

Visit official page →

About

CogVideoX-5B is the 5-billion-parameter text-to-video model from Zhipu AI and Tsinghua University (THUDM), released August 2024 as the flagship of the CogVideoX family. It uses a 3D VAE plus an "expert transformer" architecture with progressive training (mixed-duration and resolution-progressive) to generate coherent 6-second videos at 720×480, 8fps, from English text prompts.

Benchmarks (all provider-reported, from the CogVideoX paper arXiv 2408.06072, published at ICLR 2025):

VBench metrics: Human Action 96.8, Scene 55.44, Dynamic Degree 62.22, Multiple Objects 70.95, Appearance Style 24.44, Dynamic Quality 69.5 (provider-reported)
Human evaluation vs Kling: CogVideoX-5B wins on all aspects — total score 2.74 vs 2.17 (sensory quality 0.722 vs 0.638, instruction following 0.495 vs 0.367, physics simulation 0.667 vs 0.561, cover quality 0.712 vs 0.668; provider-reported)

No independent replication of these numbers was found. Note the newer CogVideoX1.5-5B (Nov 2024) supersedes this checkpoint with 1360×768 output and 5–10s clips; this entry covers the original 5B release.

Specs

Context window

English text prompt (≤226 tokens) → 6s video at 720×480, 8fps (provider-claimed)

Pricing

Weights free under CogVideoX community LICENSE (custom; registration + usage caps reported — verify before commercial use). Hosted API via Zhipu BigModel (QingYing/CogVideoX endpoints) — pricing could not be verified against Zhipu's pricing page. Provider-claimed/unverified.

Free quota

Weights free to download and self-host; Zhipu historically offered free-tier API access (unverified current status)

API access

Self-host via Hugging Face weights (zai-org/CogVideoX-5b) + diffusers CogVideoXPipeline; hosted via Zhipu BigModel API and third-party inference

Weights

Open weights

Released

August 27, 2024

Benchmarks

VBench (provider-reported, paper): Human Action 96.8, Scene 55.44, Dynamic Degree 62.22, Multiple Objects 70.95, Appearance Style 24.44, Dynamic Quality 69.5. Human eval vs Kling: 2.74 vs 2.17 total (provider-reported). No independent verification found.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-videoshort-form videoopen video researchprompt-driven animation

Trust score — 6/10

/10 trustVerified

Scored 6/10 (verified tier, metrics-v1). Evidence: HF: zai-org/CogVideoX-5b; HF downloads 14,873 (peer-normalized); released 777 days ago. Manual review: RED FLAG: CogVideoX1.5-5B (Nov 2024) supersedes it; benchmarks provider-reported (paper ICLR 2025). License: custom CogVideoX LICENSE (code only is Apache-2.0; ~1M visits/mo cap per secondary reporting).

How the trust score is built

Full methodology →

Adoption · stars/downloads, log-scaled vs peers: 6.6/10
Maintenance · recency, releases, issue load: 4.0/10
Openness · license + docs: 6.5/10
Manual review · sanity read + red-flag check: 7.0/10

· HF: zai-org/CogVideoX-5b
· HF downloads 14,873 (peer-normalized)
· released 777 days ago
· manual review: RED FLAG: CogVideoX1.5-5B (Nov 2024) supersedes it; benchmarks provider-reported (paper ICLR 2025)