CogVideo is a text-to-video generative AI model developed by Zhipu AI, one of the earlier large-scale open-source models capable of generating multi-second video clips directly from text prompts. Released as an open-source research model, CogVideo contributed to advancing the field of AI video generation by demonstrating that autoregressive transformer architectures could be applied to video generation at meaningful durations and resolutions.

The original CogVideo model used a hierarchical autoregressive approach to generate video frame by frame conditioned on text, building on the CogView image generation architecture. It was notable for producing semantically coherent short clips that responded to natural language descriptions, even if visual quality was limited compared to later generation models. The model was significant as one of the first large-scale video generation models to be released openly, enabling academic research and community experimentation. Subsequent versions and derivatives in the CogVideo lineage improved visual quality, resolution, and motion coherence as the field developed.

CogVideo represents an important milestone in the progression from image generation to video generation, illustrating how architectural approaches proven on static images were extended to handle the additional temporal dimension of video. For practitioners tracking the development of AI video tools, understanding early models like CogVideo provides context for the architectural decisions and capability benchmarks that later, more capable production models have built upon.