Tencent Releases Wan2.2, a 14B MoE Video Model

Tencent has released Wan2.2, a powerful new text-to-video model, making it available to the public on the Hugging Face Hub. The model, officially named Wan2.2-T2V-A14B, stands out with its 14-billion-parameter scale and a sophisticated architecture, all provided under a permissive Apache 2.0 license.

An Efficient Architecture

At its core, Wan2.2 employs a Mixture-of-Experts (MoE) design. This approach allows the model to be very large in terms of total parameters while only activating a fraction of them for any given task. The result is more computationally efficient inference compared to a dense model of similar size. Its release in the popular Diffusers library format also simplifies integration for developers and researchers looking to build on the technology.

This release marks another significant contribution to the open-source AI video landscape. By providing a large, advanced model with a permissive license, Tencent is fueling the innovation happening outside of closed, proprietary systems. It provides a new foundation for developers creating applications in media, entertainment, and beyond, adding more competition and diversity to the field of generative video.