Zhipu AIs state-of-the-art open-source video model featuring 3D causal VAE and Expert Transformer architecture.
Weights and code fully open-sourced on GitHub and Hugging Face (CogVideoX-2B/5B), deployable on consumer and enterprise GPUs.
CogVideoX is a landmark open-source text-to-video and image-to-video foundation model developed by Chinese frontier AI powerhouse Zhipu AI in collaboration with Tsinghua University researchers. As proprietary commercial models like Sora and Runway Gen-3 keep their architectures closed, CogVideoX provides developers and creative studios with an accessible, high-performance open-source alternative. The model introduces two pivotal architectural innovations: first, a custom 3D Causal Variational Autoencoder (3D Causal VAE) that achieves an extraordinary 8x8x4 compression across both spatial and temporal dimensions, dramatically reducing memory bandwidth bottlenecks; second, a multimodal Expert Transformer backbone that unifies text tokens and video temporal latents without requiring separate cross-attention layers. Released in both 2B (accessible on consumer GPUs) and 5B parameters, CogVideoX produces photorealistic cinematic footage with smooth physical inertia, complex motion dynamics, and rich temporal coherence.